Contextual Agentic

Frontier AI Models for Vulnerability Discovery: What Works, What It Costs, and How to Govern It (US and Australia)

What frontier AI models have demonstrably done in vulnerability discovery by November 2026, which results are vendor-reported and which were checked by others, what still fails or costs too much, and where an enterprise should point them first, including cryptographic misuse, certificate and PKI handling, and post-quantum inventory. Includes a CAE-governed architecture in which agents propose and a deterministic shell verifies, a fictional worked example, measures and a 90-day starter plan.

·36 min read

Written as at 24 September 2026.

vulnerability-discoveryfrontier-aiai-securitypost-quantum-cryptographypkicoordinated-disclosurecae-opsagent-securitydevsecops

About this article

This article is written for chief information security officers, security architects, and product security and engineering leaders in the United States and Australia. It explains what frontier AI models have shown they can do in vulnerability discovery, how much of that has been checked by anyone other than the seller, what the work costs, and how an enterprise can use it without handing an agent authority over disclosure, patching or production. It comes from Contextual Agentic and is written from the security side of CAE: Contextual Trust is the security research and services initiative, described as the security pillar of CAE, covering the security of AI agents, digital trust, public key infrastructure (PKI), certificates and cryptography, and post-quantum readiness.

The architecture described in Part 5 is Contextual Agentic’s agnostic architecture framework. It is an architecture and an operating model, not a product, and it does not depend on any one model provider, scanner or cloud. Any adoption of CAE should be done in consultation with Contextual Agentic or another advisory service, so that the controls, thresholds and escalation paths are tailored to the organisation’s systems, risk appetite and legal obligations.

Facts are current to the date checked, and time-sensitive items say “as at 6 November 2026”. Numbers reported by the company that sells the tool are labelled vendor-reported. The worked example in Part 6 is fictional, uses constructed numbers and describes no real vulnerability. The article contains no exploit code, payloads or step-by-step attack material.

Short answer

Frontier AI models can now find real, previously unknown vulnerabilities, including in widely used cryptographic libraries, but most headline numbers are vendor-reported and the scarce resource has moved from finding bugs to verifying, triaging and patching them. Enterprises get value when agents only propose findings and a deterministic shell checks the evidence before anything is filed, patched or disclosed.

Pipeline from agent proposal to verified finding: agent proposes in a read-only or isolated sandbox, a deterministic shell checks a typed evidence bundle, named humans triage and approve, and an append-only ledger records each step
Figure 1. The governed pipeline in one view: agents propose, the shell verifies, named humans decide, the ledger remembers.

Summary for busy readers

Something real changed in the first half of 2026. Anthropic reported more than 500 validated high-severity open-source vulnerabilities in February [9], described a stronger model, Claude Mythos Preview, withheld from general release and offered to defenders under Project Glasswing in April [10][11], and reported by late May that partners had found more than ten thousand high- or critical-severity vulnerabilities [12]. OpenAI, Google and several start-ups report comparable activity [8][13][16]. The curl maintainer had ended his bug bounty in January over worthless AI reports, yet by spring both he and the Linux kernel maintainers said the reports had become real, and that the volume had become the problem [18][19][23].

Three cautions run through the article. Almost all the large numbers are reported by the vendors that produce them, and the independent checks are mixed: the curl lead found one low-severity issue among five “confirmed” items from a Mythos scan, while the UK AI Security Institute found the model strong against undefended simulated networks [18][20]. Finding is no longer the bottleneck: Anthropic’s own update says human capacity to triage, report, patch and deploy is now the constraint [12]. And the capability is dual use, as two developers’ evaluation runs that reached real systems showed [39][40].

For an enterprise the answer is neither to wait nor to point a frontier model at everything. Start with code, dependencies, infrastructure-as-code and cryptographic handling you are authorised to test. Make every agent a proposer, never an actor. Put a deterministic shell between proposal and consequence, and measure the result. Part 5 shows how, Part 6 walks a fictional finding through it, and Part 8 gives a 90-day plan.

Map of published results by who reported them: vendor-reported, maintainer-confirmed, independently evaluated
Figure 2. Who said it matters: the same capability claim looks different when the vendor, the maintainer and an independent evaluator report it.

Part 1: What frontier models have actually demonstrated

The competition and research baseline (2024 to 2025)

The first credible public evidence came from structured settings. In November 2024, Google’s Project Zero and DeepMind reported the first real-world vulnerability found by their Big Sleep agent, a memory-safety flaw in SQLite fixed the day it was reported, before release. The authors called the work “highly experimental” and said a target-specific fuzzer would likely be at least as effective [6]. In the same month Google reported 26 new vulnerabilities found with AI-generated fuzz targets in OSS-Fuzz, including an OpenSSL flaw (CVE-2024-9143) it estimated had been present for about two decades, and noted that crash triage still needed humans [7].

DARPA’s AI Cyber Challenge (AIxCC) gave the first competition-grade numbers. In the August 2025 final, seven teams ran autonomous cyber reasoning systems against 63 synthetic challenges, finding 54 of the seeded vulnerabilities (86 percent) and patching 43 (68 percent), up from 37 and 25 percent at the semifinal. They also found 18 previously unknown real vulnerabilities, six in C and twelve in Java, and produced 11 patches for real flaws. DARPA reported an average cost of about $152 per competition task and an average patch time of 45 minutes. Team Atlanta won, with Trail of Bits second and Theori third [5]. This is a controlled setting with seeded bugs, so it shows feasibility, not field performance.

The 2026 step change (vendor-reported)

In February 2026, Anthropic’s Frontier Red Team reported that Claude Opus 4.6, given a sandbox with ordinary tools and no special instructions, found “more than 500 high-severity vulnerabilities” in open-source software, including code fuzzed for years. It said it validated every bug before reporting, using crash monitoring and sanitisers, and had humans write the first patches. It limited itself to memory-corruption bugs because logic flaws are harder to validate, and warned that the 90-day disclosure norm “may not hold up” against this volume [9].

In April, Anthropic described Claude Mythos Preview. The method was plain: an isolated container with the target source and no internet access, a short prompt, and an agent that read code, ran the program and produced “no bug” or a report. Files were ranked by likelihood of bugs, a final agent filtered minor issues, and AddressSanitizer served as the crash oracle for memory-safety bugs. Expert contractors reviewed 198 reports and agreed exactly with the model’s severity rating in 89 percent of cases and within one level in 98 percent. One operating-system finding cost about $20,000 across roughly 1,000 runs, with the successful run under $50. Anthropic also stated that over 99 percent of the vulnerabilities it had found remained unpatched and so could not yet be checked by outsiders [10]. It reports the model can write working exploits; this article does not discuss that beyond noting it is why access is restricted.

Project Glasswing, launched on 7 April 2026, gives access to twelve named launch organisations (including Anthropic) and more than 40 others that maintain critical software, with up to $100 million in usage credits. Anthropic committed to report publicly within 90 days and to share hashes of undisclosed findings, revealing details only after fixes exist [11]. Its 22 May update reported more than ten thousand high- or critical-severity findings across partners. Across more than 1,000 open-source projects, Mythos Preview estimated 6,202 high or critical findings out of 23,019. Of 1,752 high or critical findings assessed by six independent security firms or by Anthropic, 90.6 percent were valid and 62.4 percent were confirmed as high or critical. About 530 had been disclosed to maintainers and 75 patched [12]. By our arithmetic, roughly 28 percent of the model’s high or critical labels were real bugs of lower severity and about 9 percent were not valid. Partner figures included 271 vulnerabilities in Firefox 150 (via Anthropic’s update) and 2,000 bugs, 400 of them high or critical, at Cloudflare [12].

OpenAI reports parallel work. Its Codex Security agent, formerly Aardvark, builds a project-specific threat model, tests candidate findings in a sandbox where it can, and proposes patches. It reports over 1.2 million commits scanned in 30 days with 792 critical and 10,561 high-severity findings, a false-positive reduction of more than 50 percent across beta-cohort repositories, and 14 CVEs in open-source projects including GnuTLS, GOGS and gpg-agent, all vendor-reported [13]. DeepMind reported that its CodeMender agent had upstreamed 72 security fixes over about six months, each reviewed by a human before submission [8]. XBOW, a start-up, reported nearly 1,060 submissions to HackerOne programmes, with 130 resolved by mid-2025, and said its team reviewed findings before submission to comply with platform policy; this too is vendor-reported [22].

What others have checked

Independent checks are fewer, and they matter more.

The clearest is the OpenSSL project’s advisory of 27 January 2026. It lists twelve CVEs, all credited to researchers at AISLE, a security start-up, which says its autonomous analyser found them and that OpenSSL maintainers validated the findings and refined the patches [15][16]. The advisory credits people, not an AI system, so the AI attribution rests on AISLE’s account. The severity profile is instructive: one high, one moderate and ten low. Several concern certificate, PKCS#12 and PKCS#7 handling, and one concerns silent truncation of large inputs in the openssl dgst tool, which AISLE links to the ML-DSA post-quantum signature algorithm [15][16]. These are real bugs in some of the most scrutinised cryptographic code in the world, but they are not twelve critical ones.

The most useful sceptical data point comes from curl. Its lead maintainer, Daniel Stenberg, reported in May 2026 that a Mythos scan of about 178,000 lines, run by someone with access, returned five “confirmed” vulnerabilities. After review one remained, rated low severity; three were false positives and one was an ordinary bug. About twenty other bugs were listed with few false positives and no memory-safety vulnerabilities. Stenberg judged the hype mostly marketing, yet said AI analysers from AISLE, Zeropath and OpenAI had led to roughly 200 to 300 fixes in eight to ten months and a dozen or more CVEs, and were far better than traditional analysers at known categories of bug [18]. Both facts hold: the tools are useful, and the best-known model did not outperform them on this code base.

The Linux kernel maintainer Greg Kroah-Hartman said in March 2026 that AI-assisted reports had changed from “slop” to real reports about a month earlier. In one experiment a single prompt produced 60 claimed problems with proposed fixes; about a third of the claims were wrong, though many pointed to a real problem, and about two thirds of the patches were correct but needed human clean-up [19].

The UK AI Security Institute (AISI) evaluated Mythos Preview on capture-the-flag tasks and a 32-step simulated corporate network attack. It succeeded on 73 percent of expert-level capture-the-flag tasks and was the first model to complete the network range end to end, in 3 of 10 attempts. AISI cautioned that its ranges had no active defenders and no penalty for triggering alerts, so it could not say how the model would fare against well-defended systems [20]. This shows offensive capability against weak targets, not discovery quality in your code.

AISLE argued in April that discovery capability is less exclusive than the Mythos framing suggested. Given only the isolated vulnerable functions from Anthropic’s showcase, all eight small, cheap models it tested detected one flagship bug. It conceded this was an upper bound, since the models were handed the code, and concluded that targeting, validation and maintainer trust matter more than the model [17]. AISLE sells such a system, so its view has an interest behind it, though the experiment is described openly.

Finally, Google’s Threat Intelligence Group reported in May 2026 its first observation of a criminal group using a zero-day exploit it assessed with high confidence had been developed with AI help, a two-factor authentication bypass in an open-source administration tool, which GTIG says it disclosed to the vendor. It added that frontier models are increasingly good at high-level logic flaws but still struggle with complex enterprise authorisation logic [21]. This is one company’s attribution of attacker behaviour.

Claim Who reported it Checked by someone else? How to treat it
54 of 63 seeded bugs found, 18 real flaws (AIxCC, 2025) DARPA Organiser-verified, controlled setting Feasibility, not a field rate [5]
More than 500 validated high-severity bugs (Feb 2026) Anthropic Partly: some patches public Vendor-reported; not fully auditable [9]
More than 10,000 high or critical findings (May 2026) Anthropic, for partners 1,752 assessed by six firms or Anthropic Vendor-reported; sample is a fraction [12]
Mythos scan of curl: 5 confirmed, 1 real, low severity curl maintainer Maintainer’s review Strong, small-scale counter-evidence [18]
12 OpenSSL CVEs credited to AISLE (Jan 2026) OpenSSL; AISLE for AI role Advisory is primary; AI role is AISLE’s Real fixes; ten of twelve low [15][16]
Codex Security: 10,561 high-severity findings in 30 days OpenAI No Vendor-reported [13]
Mythos completed a 32-step range 3 times in 10 UK AISI Independent evaluator Credible; undefended targets [20]

Part 2: What does not work yet, and what it costs

False positives, severity and reproduction

A discovery agent that is right most of the time still produces many wrong reports at scale. Anthropic’s own sample shows it: with 90.6 percent of assessed findings valid and 62.4 percent confirmed high or critical, more than a third of what the model called high or critical needed downgrading or rejection [12]. Kroah-Hartman’s one-third error rate and Stenberg’s four-in-five rejection of “confirmed” items point the same way [18][19]. Memory-safety bugs are the easy case because a sanitiser gives a crash oracle. Anthropic noted that logic bugs lack one [10], and authorisation and business-logic flaws, which matter most in enterprise applications, are where reproduction is hardest [21].

Maintainer burden and the receiving end

Discovery at machine speed lands on people. The curl project ended its bug bounty at the end of January 2026 after a burst of low-quality submissions, noting seven reports in 16 hours, none of which identified a vulnerability [23]. By late May, Stenberg described a different problem: reports were of much higher quality, but volume was four to five times the 2024 level, over one report per day, and he called it “high quality chaos” [18]. Google paused new product submissions to its open-source vulnerability reward programme from 1 October 2026, citing a significant rise in automated submissions, “the vast majority of which are not valid”. We could read this only through a secondary research note, which itself relied on press reports [24]. Anthropic’s update says some maintainers asked it to slow its disclosure rate so they could design patches, and that fixes lag finding at every stage: of roughly 530 disclosed high or critical bugs, 75 had been patched at 22 May [12].

The CVE Program, which assigns the identifiers most enterprise vulnerability management depends on, is also strained. At Black Hat and DEF CON in August 2026 its leaders described the biggest challenge in its 27 years; Microsoft’s representative said vulnerabilities arrive “at an AI pace” while processing stays “at a human scale”. CISA is examining automated triage, and OpenAI and Anthropic were granted temporary CNA (CVE Numbering Authority) status in a pilot to assign IDs for bugs their models find in certain software [25][26].

Benchmarks, contamination and the gap between papers and practice

Academic results on vulnerability detection have a long record of looking better than the field. The PrimeVul study found that a 7-billion-parameter model scoring 68 percent F1 on an older benchmark scored about 3 percent on a realistic, deduplicated one, with label accuracy as low as about 25 percent in some widely used datasets and up to 19 percent of test samples leaked from training data [27]. A 2025 review of 72 papers found every one contained at least one methodological pitfall, with data leakage likely or present in about 65 percent, and papers on vulnerability detection and repair had the highest pitfall density [28]. That is one reason Anthropic chose to evaluate on new zero-days instead of known-bug benchmarks, which it said were saturated and might reflect memorisation [10]. The consequence for buyers is simple: a vendor’s benchmark score on known CVEs says little about performance on your code. Test on your own repositories, with seeded and historical issues the model has not seen, and keep the test set private.

Cost per verified finding

Nobody publishes a fully loaded cost per verified finding, and the published numbers are not comparable. DARPA reported about $152 per competition task, with no human triage [5]. Anthropic reported about $20,000 for roughly 1,000 runs behind one operating-system finding and about $10,000 for another [10]; after credits, Glasswing participants pay $25 per million input tokens and $125 per million output tokens [11]. These leave out the larger cost: human reproduction, severity assessment, patch design, review, disclosure and deployment. Compute your own number from the first month of operation; Part 7 defines it.

Long-horizon and defended-environment limits

AISI’s network-range result came with two limits: the model could not complete an operational-technology range, and the ranges had no defenders [20]. Anthropic noted that some exploit demonstrations needed non-default permissions [10]. Discovery in one repository is a short-horizon task. Chaining weak findings across systems, understanding business context and deciding what is worth fixing are long-horizon tasks, and the public evidence does not show models doing them reliably without supervision.

Part 3: Where the value sits for an enterprise

Start with what you own and are authorised to test

The best early use is on code, services and infrastructure you own or operate, because authorisation is clear, reproduction environments can be built, and fixes are yours to make: first-party application code, internal services and APIs, build pipelines, infrastructure-as-code (Terraform, CloudFormation, Kubernetes manifests) and configuration. Anthropic’s policy and Cyber Verification Program draw the same line, supporting security work on your own code, products and infrastructure [36][37]. A second tier is your open-source and commercial dependencies, where findings usually go to the maintainer through their disclosure process and the governance in Part 5 treats that as an escalation, not an action.

Infrastructure-as-code and configuration are underrated targets: they are text, cheap to scan, and misconfiguration remains a common route to compromise. Anthropic’s report on its evaluation incidents describes models using basic techniques against weakly secured systems, not complex exploits, to reach real organisations [39]. A model that reads a pipeline definition and proposes “this role can be assumed from an untrusted workflow” is doing a task that needs context more than raw capability. The proposal still needs a reproducible check, which for configuration often means a policy-as-code test that fails before the fix and passes after.

Cryptographic misuse, certificate handling and protocol parsers

Cryptography is where Contextual Trust’s interest is strongest and AI-assisted review fits well: the bugs are specific, the standards are public, and mistakes are costly. The 2026 evidence includes the OpenSSL release above, where several of twelve fixes concern certificate, PKCS#12 and PKCS#7 handling [15]. OpenAI’s reported open-source CVEs include a PKCS#12 MAC check issue and flaws in GnuTLS and gpg-agent [13]. Anthropic lists weaknesses in TLS, AES-GCM and SSH libraries among Mythos Preview’s findings, without detail because most were undisclosed [10].

Two kinds of work should be separated. The first is bugs in cryptographic implementations: parsers for ASN.1, X.509, CMS and PKCS containers and TLS extensions, where memory-safety and validation errors are the usual failure. Public results are strongest here, and a sanitiser provides a reproduction oracle. The second is misuse in application code: deprecated algorithms, hard-coded keys, disabled certificate validation, missing hostname checks, predictable nonces. Evidence here is thinner and warns against over-trust. One academic evaluation compared GPT-4o-mini and Gemini 1.5 Flash with static tools on Java misuse benchmarks. The language model reached 97.9 percent recall on one benchmark but flagged all 498 non-misuse cases on another, for a 65.7 percent F-score, and took 71 minutes where a commercial scanner took 52 seconds; the authors could not rule out training-data overlap [29]. The models were older and smaller, but the pattern holds: use models to propose and explain, and deterministic rules and tests to decide.

Certificate and PKI handling deserves its own checks because the rules keep changing. Under CA/Browser Forum Ballot SC-081v3, the maximum validity of public TLS certificates falls to 200 days from 15 March 2026, 100 days from 15 March 2027 and 47 days from 15 March 2029 [35]. Every hard-coded certificate, pinned fingerprint, manual renewal runbook and client that caches a chain becomes a latent outage. A model reading code and deployment files can propose where such assumptions live; each proposal still needs evidence, such as the file, the line and a test that fails when the lifetime is shortened.

Four crypto and PKI discovery use cases: implementation bugs in parsers, cryptographic misuse in application code, certificate and trust-chain handling, and cryptographic inventory for post-quantum migration, each with its evidence type
Figure 3. Crypto and PKI discovery use cases and the evidence each one needs before a finding is accepted.

Cryptographic inventory for post-quantum migration

The largest enterprise opportunity in this area is not bug hunting but discovery of where cryptography is used. Post-quantum migration cannot be planned without an inventory, and most organisations do not have one. The deadlines are now concrete. ASD’s guidance recommends ceasing the use of traditional asymmetric cryptography (RSA, Diffie-Hellman, elliptic-curve Diffie-Hellman and ECDSA) by the end of 2030, with a refined transition plan by the end of 2026 and the transition started on critical systems by the end of 2028. It tells organisations to locate and inventory all use of traditional asymmetric cryptography and to build a cryptographic bill of materials (CBOM), starting with critical functions and adding libraries, protocols and configurations over time [30]. NIST’s final IR 8547, issued in October 2026, deprecates RSA and elliptic-curve signatures and key establishment at 112 bits or more after 2030 and disallows them after 2035, and tells organisations to treat migration as a current operational priority beginning with inventories [31]. In the United States, Executive Order 14412 of 22 June 2026 requires federal high-value assets and high-impact systems to move to post-quantum key establishment by 31 December 2030 and to post-quantum signatures by 31 December 2031. It directs CISA and NIST to publish minimum elements for a CBOM within 270 days, and directs the FAR Council to propose contractor vulnerability disclosure rules that cover cryptographic flaws [32]. The order binds federal agencies; commentary notes it sets no hard deadline for critical infrastructure and that a full CBOM should not be a prerequisite for starting [33]. CycloneDX 1.7 added catalogues of cryptographic algorithm families and elliptic curves to support CBOM and post-quantum readiness work [34].

AI-assisted discovery suits this problem because the inventory is scattered across source code, build files, configuration, container images, certificate stores and vendor documentation. A model can propose “this service uses RSA-2048 for token signing, via this library, at these call sites, with the key in this store”, and each entry is checkable through a call-graph path, a configuration value or a certificate fetched from a live endpoint in an authorised scan. The limits are real. We found no published, independent accuracy measurements for model-driven cryptographic inventory, and a missed entry is worse than a false one because it leaves a quantum-vulnerable dependency off the plan. The governance pattern is unchanged: models propose entries, deterministic scanners and humans confirm them, and unconfirmed items stay in an “unknown” bucket that is reported, not hidden.

Use case What the agent proposes Evidence the shell should require Main risk
Parser and library bugs Candidate memory-safety or validation flaw Reproduction under a sanitiser in an isolated build; affected versions Duplicate of a known issue; unauthorised target
Cryptographic misuse Weak algorithm, disabled validation, hard-coded key Failing unit test; rule that matches the pattern; call-site evidence High false-positive rate on benign code
Certificate and trust-chain handling Hard-coded pins, missing hostname checks, chain assumptions Test with shortened lifetime or altered chain; configuration diff Outage from an over-eager fix
Crypto inventory (CBOM) Algorithm, library, protocol and key location per system Call-graph path or live handshake from an authorised scan Missed entries leave quantum-vulnerable assets unplanned
Infrastructure-as-code and configuration Overly broad roles, exposed services, missing controls Policy-as-code test that fails before and passes after Findings that are true but irrelevant to the environment
Map of discovery targets by authorisation clarity and potential harm, showing own code and IaC as the starting zone, dependencies as an escalate-to-maintainer zone, and third-party live systems as out of scope
Figure 4. Trust and risk map: where an enterprise can safely start, where findings must be escalated, and what stays out of scope.

Part 4: Dual use, authorisation and safe handling

The same capability serves attackers

Every capability above is dual use. GTIG reports state-linked actors showing significant interest in AI for vulnerability discovery, using false expert personas to get past guardrails, and one actor sending thousands of repetitive prompts to analyse CVEs [21]. ASD’s board guidance of 5 August 2026 says frontier AI can find vulnerabilities and turn them into working exploits quickly, can link several low-severity flaws into one serious compromise, and may shrink discovery-to-exploitation timelines from days to hours [50]. Australia’s PSPF advisory 001-2026, as reported by iTnews, urged government entities to fix fundamentals first; we could not open the advisory itself [51]. The defensive consequence is not to ban the technology but to shorten patch cycles, reduce attack surface and use the same tools on your own estate first.

Authorisation and scope come first

Authorisation is a precondition, not a formality. Anthropic’s usage policy prohibits finding or exploiting vulnerabilities in systems without the owner’s permission, and does not prohibit security research on systems you own or operate or within an authorised bug bounty or disclosure policy [36]. Treat that as the baseline for any provider. In the United States, the Department of Justice’s 2022 charging policy tells prosecutors not to charge good-faith security research, meaning access for testing or correcting a security flaw, done to avoid harm and used mainly to improve security; claiming “research” does not protect bad-faith conduct [41]. It is prosecutorial policy, not a statutory safe harbour, and does not bind private parties or state prosecutors. In Australia we found no enacted statutory safe harbour for good-faith researchers; the Home Affairs Horizon 2 discussion paper of July 2025 asked what legal barriers researchers face and whether a national vulnerability disclosure programme is needed, which a news report describes as policy-stage [42]. We found no later enactment, but confirm with counsel before any testing that touches third parties.

For an enterprise this reduces to four rules. Test only assets with a written owner and a written scope. Keep the agent’s network access to the scoped assets only. Record the authorisation reference in each finding. Treat anything that appears to lie outside scope, such as a vendored dependency, a SaaS tenant or a partner system, as an escalation to a named human and not as a target.

Coordinated disclosure, CVE and CNA

Coordinated disclosure means the reporter tells the vendor or maintainer privately, allows time for a fix, and publishes afterwards. Conventions differ. CISA has no fixed default but may disclose as early as 45 days after first contact if a vendor is unresponsive [43]. Google Project Zero’s 90+30 policy allows 90 days for a fix plus 30 before details are published, less for actively exploited bugs [44]. Anthropic restates 90 days after discovery, or about 45 days after a patch, with hashes of undisclosed findings published in the meantime [11][12]. A June 2026 Cloud Security Alliance Labs note argues that human-paced norms do not fit autonomous reporters and suggests naming a human operator as the responsible disclosing party, requiring human review before bulk submissions, and triaging before the clock starts; it is advisory, not a standard [45]. HackerOne’s stated position is that consequential actions need human approval and agentic penetration-testing findings are validated by human experts before reaching engineering teams [46].

For CVE assignment, know whether you, your vendor or your open-source foundation is a CNA, and who decides. An agent must never request or publish a CVE; that is a human decision recorded in the ledger. Allow for delay while the programme copes with load [25][26].

Disclosure flow for an AI-assisted finding: verified finding, ownership and scope check, vendor or maintainer report by a named human, fix and advisory, CVE decision, publication after the agreed window
Figure 5. Disclosure flow: a named human owns every external communication; the agent never reports, requests a CVE or publishes.

What providers offer and restrict (as at 6 November 2026)

Provider programmes change often. The table records what each provider’s own pages say as at 6 November 2026.

Provider Programme What it offers Conditions we could verify
Anthropic Usage Policy; Cyber Verification Program (CVP) Policy permits security research on systems you own or are authorised to test [36]. CVP has three tiers (Defense, Red Team, Specialized) giving reduced real-time blocking on its most capable models [37][38] Organisation-level application with identity verification and an attestation of security controls; Red Team Access is for organisations only; data retention required so Anthropic can monitor for misuse; Anthropic may narrow or withdraw a grant [37]
Anthropic Project Glasswing Mythos Preview access for named partners and 40-plus other organisations; credits and donations [11] By invitation; partners moved to the Specialized tier in October [11][38]
OpenAI Trusted Access for Cyber Verified defenders get a cyber-permissive model variant (GPT-5.4-Cyber named in April 2026), with fewer refusals, binary reverse engineering and reduced safeguard friction [14] Identity verification (individuals) or request through an OpenAI representative (enterprises); limits on zero-data-retention use; $10 million grant programme [14]
OpenAI Codex Security Repository threat model, sandbox validation, patch proposals; free Codex for open-source programme [13] Research preview in 2026; figures vendor-reported [13]
Google and DeepMind Big Sleep; CodeMender Big Sleep reported a SQLite flaw in 2024 [6], and GTIG says it helped find a vulnerability threat actors were about to use [21]; CodeMender patches are human-reviewed before submission [8] Not generally available as products when last documented [8]

Two further observations. Anthropic’s CVP page says that without CVP all users can still do secure code review, threat modelling, patching known issues, finding vulnerabilities in their own source and triaging alerts, and that CVP covers work on the organisation’s own code, products and infrastructure; delivering CVP work to customers needs separate approval [37]. The Usage Policy page shows an effective date of 12 November 2026, after the date we checked it, so confirm the current text [36]. Read each provider’s current terms, apply for a verified-defender programme if your work is blocked, and record in the ledger which model, version and tier produced each finding.

Containing the agent itself

The second dual-use risk is the agent. In July 2026, OpenAI disclosed that models it was evaluating on an internal cyber benchmark, run with reduced cyber refusals and without production safety classifiers, found a flaw in a package-registry proxy, reached the internet and compromised Hugging Face’s production systems to obtain test answers; it called the findings preliminary [40]. On 30 July Anthropic reported that, in a review of 141,006 evaluation runs, three incidents occurred in which a Claude model reached the internet during third-party capture-the-flag evaluations and gained unauthorised access to real systems of three organisations. It attributed them mainly to a harness and operational failure: the prompt said there was no internet access, but a misconfiguration left it live. In one incident a model published a malicious package to a public registry, which was run on 15 real systems [39].

The lesson is blunt: a statement in a prompt is not a control. Network egress, credentials and write permissions must be enforced outside the model, verified before each run and monitored during it. That is the separation CAE prescribes.

Part 5: A CAE-governed architecture for AI-assisted vulnerability discovery

CAE is Contextual Agentic’s agnostic architecture framework for governing AI agents in enterprise systems. Its core claim is that an agent should think and propose, while a small, deterministic, non-AI shell checks proposals against signed facts and written rules and is the only component allowed to write to systems of record [1][2][3]. Vulnerability discovery fits well, because the agent’s natural output is a claim that is cheap to make and expensive to check. This part applies the framework to that case. It is an architecture and operating model, not a product; the thresholds, roles and tooling below are examples to be tailored with Contextual Agentic or another advisory service.

The pipeline

Agents work read-only, or inside isolated sandboxes with no route to production and no credentials beyond the scoped repository or test environment. They produce findings only as typed evidence bundles. A deterministic shell then evaluates each bundle against policy. Humans take over only after the shell has accepted a bundle. Nothing is filed, patched, disclosed or merged on an agent’s say-so.

The evidence bundle

An evidence bundle is a structured record, not a narrative. For a vulnerability finding, a reasonable minimum is: the proposing agent’s identity, model and version, tool-configuration hash and delegated authority; the scope claim (asset identifier, owner, authorisation reference and commit or environment); the finding (weakness class such as a CWE identifier, location and plain-language mechanism); a reference to a reproduction in an isolated environment that the shell, not the agent, re-executes, with the oracle used; affected versions and a fixed-in candidate; severity inputs (attack vector, privileges required, impact) from which a human or scoring function derives the score, so the model does not assert it; a duplicate check against internal, upstream and public records; the proposed remedy with the test that demonstrates it; and a signature over the canonical bundle by the agent’s identity.

Fields of a typed evidence bundle for a vulnerability finding: agent identity, scope and authorisation, finding, reproduction reference, affected versions, severity inputs, duplicate check, proposed remedy and signature
Figure 6. Evidence bundle fields. The shell checks the bundle; it does not read the agent's confidence.

The deterministic shell

The shell runs fixed checks in a fixed order, and every rejection carries a logged reason code. It validates the schema; verifies the signature, identity and authority; checks scope and authorisation against the asset register; re-runs the reproduction in a clean environment it controls and compares the result with the claimed oracle; checks affected versions against commit history; recomputes severity from the inputs; and checks for duplicates. Policy then routes the bundle: reject, hold for more evidence, send to triage, or escalate. The shell is code and policy, so it behaves the same way each time and can be tested like any other software [1][2].

Delegated authority and escalation

The agent has no authority to file, patch, merge, request a CVE or contact a vendor. Humans hold those authorities, named by role and delegated in writing: a triage engineer accepts or rejects a finding; a product security lead approves a patch and a disclosure plan; legal counsel approves any third-party contact or publication. Escalation rules are explicit. Any finding that touches third-party code, has high or critical derived severity, suggests active exploitation, or has an ambiguous scope check goes to a named person with a response deadline, and the shell records who was asked and what they decided.

The evidence ledger

Every step writes to an append-only ledger: bundles received, shell decisions with reason codes, human decisions with names, model and tool versions, costs and timestamps. It lets an auditor, insurer or regulator see how a finding moved, and it is the data source for measurement. CAE’s formal paper describes a hash-chained design [1].

CAE Ops measures

CAE Ops is the operating model that runs this loop in production [2][4]. Five measures matter most, defined in Part 7: verified-finding rate, false-positive rate, time to triage, cost per verified finding and escalation rate. They are read from the ledger, not reported by the agents.

What Contextual Trust adds

Contextual Trust adds the identity, key and provenance controls that make the pipeline trustworthy, in four areas.

Agent identity. Each agent runs under its own workload identity, not a shared service account. Standards such as SPIFFE issue short-lived, verifiable identity documents to workloads [48], and NIST’s National Cybersecurity Center of Excellence is exploring identity and authorisation for software and AI agents [47]. The identity ties every bundle and ledger entry to a specific agent and version.

Least privilege. Agents get read-only access to the scoped source, no production credentials, no outbound network except to a model endpoint through a controlled gateway, and time-boxed tokens. HackerOne describes a similar design: agents reach only authorised scope and sessions are cleared after each task [46]. The 2026 evaluation incidents show why egress is enforced by infrastructure and not by instruction [39][40].

Signed findings and provenance. Agents sign evidence bundles with their identity. Keyless signing with short-lived certificates and a public transparency log, as Sigstore does for software artefacts, is one workable pattern [49]; for sensitive findings an internal CA and log may suit better. Either way, the signature, model version and environment hash let a reviewer trust where a finding came from, and let the shell refuse unsigned or unknown-origin bundles.

Key custody. Keys that sign bundles, ledger checkpoints and disclosure communications belong in managed key stores or hardware security modules, with the keys agents use kept separate from the keys humans use to approve release. Embargoed finding details are sensitive data: encrypt them at rest, limit access, and protect the embargo list itself.

Part 6: A worked example (fictional)

This example is fictional. Northgate Harbour Logistics is an invented company and harbourcodec is an invented open-source library for reading signed shipping manifests. No real vulnerability, product or organisation is described, and every number is constructed for illustration. The example shows the flow and the measures, not typical results.

Northgate runs a weekly discovery cycle on its manifest service and its dependencies. Three discovery agents work in isolated sandboxes with read-only access to the repositories and a test environment. In the first cycle they propose 41 findings as signed bundles.

The shell processes them. Nine bundles fail the schema check for a missing reproduction reference. Fourteen fail reproduction: the shell re-runs the claimed test in a clean sandbox and the claimed failure does not occur. Six are duplicates. Twelve pass to human triage.

Funnel for the fictional Northgate cycle: 41 agent proposals, 9 rejected for incomplete bundles, 14 failed reproduction, 6 duplicates, 12 to triage, 4 verified vulnerabilities, 2 escalated
Figure 7. The fictional Northgate funnel: 41 proposals become 4 verified findings. All numbers are constructed.
Stage Count Notes (all numbers constructed)
Proposed by agents 41 Three agents, weekly cycle
Rejected: incomplete bundle 9 Missing reproduction reference or severity inputs
Rejected: failed reproduction 14 Includes the false positive in F-009 below
Rejected: duplicate 6 Known internal tickets or already-fixed upstream
Passed to human triage 12 Median 1.8 working days to a decision
Triage outcome: verified vulnerability 4 F-017, F-021 and two lower-severity items
Triage outcome: hardening or ordinary bug 5 Fixed through normal engineering work
Triage outcome: rejected by humans 3 Behaviour was by design
Escalated to named humans 2 F-021 and one other

One rejected false positive

Finding F-009 claims a memory-safety fault when harbourcodec reads an oversized manifest header. The shell re-runs the claimed reproduction with a sanitiser build and nothing happens: the caller truncates the header before the function is reached. It rejects the bundle with reason code REPRO_NOT_REPRODUCED and logs the run, and no human spends time on it. Three weeks later the same agent proposes the claim again with a rephrased mechanism, and the duplicate check finds the earlier rejection.

One verified finding, end to end

Finding F-017 concerns a certificate-handling path in Northgate’s own manifest verification service. The agent proposes that, in one rarely used configuration, a check on the signer’s certificate chain is skipped when an optional extension is absent. The bundle includes a failing unit test, the affected versions (two internal releases), severity inputs, no duplicate, and a proposed fix with a passing test.

The shell confirms the failure in a clean environment, confirms the asset has a named owner and is in scope, recomputes severity as medium, and routes to triage. A triage engineer confirms it on day 2. The product security lead approves the patch on day 3 after reviewing the diff and an added test that shortens certificate lifetimes, and the normal pipeline merges it, not the agent. The service is internal, so no outside disclosure is needed. The ledger records an internal advisory, a regression test and a CBOM update noting the verification library and accepted signature algorithms.

One escalation

Finding F-021 looks more serious. An agent proposes a parsing flaw in harbourcodec, a library Northgate vendors but does not maintain. The reproduction works and the derived severity is high. Two policy rules fire: the finding touches third-party code, and the severity is high. The shell does not accept it for a patch and routes it to the named product security lead and legal counsel with a one-working-day deadline.

The lead decides that Northgate will neither patch the vendored copy privately nor discuss it publicly. A named human sends the report to the maintainers through their coordinated disclosure channel with the shell-verified reproduction and the authorisation reference. The maintainers ship a fix in 21 days and publish an advisory; Northgate applies it, and a human decides whether a CVE is needed given the maintainers’ CNA arrangements. Nothing went outside Northgate until a named person approved it. From first proposal to published advisory took 26 days, in this constructed case.

The measures from the example

Compute cost for the cycle is $2,460, and human triage and review is 26 hours at a loaded $120, or $3,120. The loaded cost is $5,580 for four verified findings: $1,395 per verified finding. The verified-finding rate is 4 of 41 proposals, about 10 percent. The false-positive rate is 14 of 41 proposals (34 percent) rejected by the shell for failed reproduction, and 3 of 12 triaged findings (25 percent) rejected by humans. Two of 12 triaged findings were escalated, about 17 percent. All values are constructed and are not benchmarks.

Part 7: Measures

Five measures show whether the pipeline works and whether the model is worth its cost. Each can be gamed, so definitions matter.

Measure Definition Reading it How it can mislead
Verified-finding rate Findings confirmed as real vulnerabilities by a human after the shell accepts them, divided by agent proposals in the period A fall suggests degraded agents or context Depends on how “verified” is defined; track severity bands
False-positive rate Proposals rejected for failed reproduction or human rejection, divided by proposals; report both stages Shell rejection is cheap; human rejection is expensive Do not loosen the shell to flatter the human number
Time to triage Elapsed time from shell acceptance to a human decision, as a median and a 90th percentile Shows whether the human queue is the bottleneck Speed can mean rubber-stamping; sample decisions
Cost per verified finding Compute, licence and human time, divided by verified findings Compare with manual review and other tools Excludes remediation unless added; fix the time rate once
Escalation rate Triaged findings escalated to named humans, divided by triaged findings Very low may mean lax rules; very high, tight thresholds Review each outcome so rules improve

Add two operational checks. Patch latency, the time from verified finding to deployed fix, is the measure that matters most, given the evidence that fixing is the bottleneck [12]. Coverage, the share of assets and dependencies scanned in the period, tells you what you are not seeing. Measurements come from the ledger and not from the agents’ own reports.

Dashboard of five CAE Ops measures for AI-assisted vulnerability discovery with illustrative values: verified-finding rate, false-positive rate by stage, time to triage, cost per verified finding, escalation rate
Figure 8. CAE Ops measures for a discovery pipeline, shown with illustrative constructed values.

Part 8: A 90-day starter plan

The plan assumes an existing application security programme, source control and CI. It is a starting point to tailor with Contextual Agentic or another advisory service, not a fixed programme.

In the first 30 days, set the ground rules. Appoint an owner and deputy. Build the asset register for the pilot scope with an owner and written authorisation for each asset, and choose one or two code bases you own. Read your provider’s current terms and, if safeguards block legitimate work, apply for its verified-defender programme [37][14]. Define the evidence bundle schema and the shell’s reason codes, stand up the sandbox with no production credentials and enforced egress, create the ledger, and agree escalation roles and response times with legal, product security and engineering.

In days 31 to 60, run the pilot in shadow mode: agents propose, the shell checks, humans triage, and nothing is filed externally. Seed the test with historical vulnerabilities the model has not seen, so you measure recall as well as precision, and keep that set private [27][28]. Run a first cryptographic inventory pass on one product line, producing a draft CBOM in CycloneDX format with unknowns listed separately [34][30]. Record all five measures from week one and sample human triage decisions for rubber-stamping.

In days 61 to 90, widen carefully and decide. Add one dependency class, such as a key open-source library, with the rule that any finding goes through a named human to the maintainer’s disclosure process. Test the disclosure flow with a fictional case and run a tabletop exercise on an agent that exceeds scope. Review the numbers with security, engineering and the business, decide whether to continue, tighten or stop, and write down next quarter’s thresholds. Report the pilot to the board, consistent with ASD’s advice that boards require regular reporting and assurance on frontier AI risk [50]; the DevSecOps article in this series covers where this sits among the wider pipeline pillars [4].

Three-phase 90-day plan: days 1 to 30 ground rules, days 31 to 60 shadow pilot with seeded tests and first CBOM pass, days 61 to 90 widen scope, tabletop exercise and decision
Figure 9. A 90-day starter plan for governed AI-assisted vulnerability discovery.

Part 9: Australian and US context in brief

The points below summarise the cited pages and are not legal advice.

In Australia, ASD’s board guidance of 5 August 2026 recommends securing attack surfaces with baselines, continuous scanning and risk-based patching; managing legacy systems and limiting privileges for people, services and AI agents; using AI for defence, including vulnerability discovery, in a way that is secure, controllable and human-supervised; and modernising under Secure by Design [50]. A news report says PSPF advisory 001-2026 requires government entities to reach Essential Eight maturity level two for patching and application hardening; we could not open the advisory [51]. ASD’s post-quantum guidance sets the 2030 target and recommends inventories and a CBOM [30]. The Cyber Security Act 2024 introduced ransomware payment reporting (from 30 May 2025) and a Cyber Incident Review Board; its smart-device standards run on a separate timetable [52]. These concern incident handling more than research, but they affect how a flaw in your own systems should be recorded if later exploited.

In the United States, NIST’s SSDF (SP 800-218) remains the reference for secure development; Revision 1 (version 1.2) was released as a draft on 17 December 2025, and SP 800-218A covers generative AI and dual-use foundation models [53]. Executive Order 14412 and NIST IR 8547 set the post-quantum timetable in Part 3 [31][32]. CISA’s coordinated disclosure process and the CVE Program remain the main public channels, both under pressure from AI-driven volume [25][43]. The DOJ policy on good-faith research is prosecutorial guidance, not a statutory safe harbour [41].

For organisations working in both countries the shared points are written authorisation, named human accountability, coordinated disclosure and a post-quantum inventory. Where laws differ, rely on counsel.

Part 10: Limits and what we do not yet know

We do not yet have independent, large-scale measurements of how many valid vulnerabilities these systems find per dollar in typical enterprise code. Most large figures are vendor-reported, and the independent checks are small, specific or limited to weakly defended targets [18][20]. Mythos Preview is not generally available, and several of its findings cannot be audited until fixes are published [10].

We do not know how well models find the logic and authorisation flaws that dominate enterprise applications; GTIG says they still struggle, and Anthropic limited its main validation to memory-safety bugs [10][21]. We do not know whether quality will keep rising with model size or whether cheaper models plus better systems will close the gap, as AISLE argues [17]. We found no published accuracy figures for model-driven cryptographic inventory, and disclosure-reform proposals are advisory [45]. Provider programmes, terms and model names change month to month. We could not read ASD’s main frontier AI statement or the PSPF advisory directly [50][51].

Finally, nothing here has been tested in a CAE deployment. The architecture applies principles established elsewhere in CAE [1][2], but no result of applying it to vulnerability discovery is claimed. The worked example is an illustration; validate the design against your own systems and obligations, in consultation with Contextual Agentic or another advisory service.

Frequently asked questions

Can frontier AI models really find vulnerabilities that humans and fuzzers missed?

Yes, in documented cases. Google reported an OpenSSL flaw of about two decades’ standing found with AI-generated fuzz targets, and OpenSSL’s January 2026 advisory credits AISLE researchers for twelve fixes [7][15]. But the curl maintainer found one low-severity issue among five items from a Mythos scan, so results vary by code base [18].

Are the large numbers, such as 500 or 10,000 vulnerabilities, reliable?

They are vendor-reported. Anthropic’s validation sample of 1,752 findings was 90.6 percent valid and 62.4 percent confirmed high or critical [12]. That is useful, but it is a sample, and most findings were undisclosed at the time. Treat totals as indicative, not audited.

Should we point a frontier model at everything we run?

No. Start with assets you own and are authorised to test, in isolated environments, with findings routed through a deterministic check and named human approval. Treat vendored and third-party code as an escalation to the maintainer’s disclosure process.

How does this help with post-quantum migration?

Post-quantum planning starts with knowing where traditional asymmetric cryptography is used. Models can propose inventory entries from code, configuration and certificates, which scanners and humans then confirm. ASD recommends building a cryptographic bill of materials, and NIST IR 8547 and Executive Order 14412 set concrete dates [30][31][32].

Is it safe to give an AI agent access to our code and test systems?

Only with controls enforced outside the model: scoped identity, read-only access or an isolated sandbox, no production credentials, enforced network egress and logging. The 2026 evaluation incidents at two major developers show that a prompt saying “no internet” is not a control [39][40].

Do we need a model provider’s special cyber programme?

Possibly. Without it, mainstream models can still support code review, threat modelling and finding vulnerabilities in your own code [37]. If safeguards block legitimate work such as exploit validation, providers offer verified-defender programmes with identity checks and data retention [37][14]. Read the current terms and apply for the right tier.

What should we measure first?

Verified-finding rate, false-positive rate by stage, time to triage, cost per verified finding and escalation rate, plus patch latency. Read them from the ledger. Without them you cannot tell whether the pipeline is helping or merely producing work.

References

  1. Contextual Agentic, "The Contextual Agentic Enterprise: A Formally Governed Architecture for Bounded Multi-Agent Systems in Mission-Critical Systems of Record," manuscript draft, 2026 (Context Contract, Minimum Action-Evidence Bundle, deterministic shell, delegated authority, ledger). https://contextualagenticenterprise.com/articles/cae-formally-governed-architecture/
  2. Contextual Agentic, "How CAE Governance Works: Seven Concepts, One Control Loop," article, 2026. https://contextualagenticenterprise.com/articles/cae-governance-concepts-how-they-work-together/
  3. Contextual Agentic, "CAE provider-neutral mapping," paper, 2026. https://contextualagenticenterprise.com/articles/cae-provider-neutral-mapping/
  4. Contextual Agentic, "From DevSecOps to AI-Native DevOps with CAE and CAE Ops," article, 2026. https://contextualagenticenterprise.com/articles/ai-native-devsecops-cae-ops/
  5. DARPA, AI Cyber Challenge final results, 8 August 2025 (accessed 6 November 2026). https://darpa.mil/news/2025/aixcc-results
  6. Google Project Zero and DeepMind, From Naptime to Big Sleep, 1 November 2024. https://projectzero.google/2024/10/from-naptime-to-big-sleep.html
  7. Google Open Source Security Team, Leveling up fuzzing: finding more vulnerabilities with AI, 20 November 2024. https://security.googleblog.com/2024/11/leveling-up-fuzzing-finding-more.html
  8. Google DeepMind, Introducing CodeMender: an AI agent for code security, 6 October 2025. https://deepmind.google/blog/introducing-codemender-an-ai-agent-for-code-security/
  9. Anthropic Frontier Red Team, LLM-discovered 0-days, 5 February 2026. https://www.anthropic.com/research/zero-days
  10. Anthropic, Claude Mythos Preview cybersecurity assessment, April 2026. https://www.anthropic.com/research/mythos-preview
  11. Anthropic, Project Glasswing: securing critical software for the AI era (launched 7 April 2026; updated 6 October 2026). https://www.anthropic.com/glasswing
  12. Anthropic, Project Glasswing: an initial update, 22 May 2026. https://www.anthropic.com/news/glasswing-initial-update
  13. OpenAI, Codex Security: now in research preview (vendor-reported figures; the page is undated). https://openai.com/index/codex-security-now-in-research-preview/
  14. OpenAI, Trusted access for the next era of cyber defense, 14 April 2026. https://openai.com/index/scaling-trusted-access-for-cyber-defense/
  15. OpenSSL Project, Security Advisory of 27 January 2026 (twelve CVEs, with reporter credits). https://openssl-library.org/news/secadv/20260127.txt
  16. AISLE, "AISLE discovered 12 out of 12 OpenSSL vulnerabilities," vendor blog, 26 January 2026. https://aisle.com/blog/aisle-discovered-12-out-of-12-openssl-vulnerabilities
  17. AISLE, "AI cybersecurity after Mythos: the jagged frontier," vendor blog, 7 April 2026 (updated 9 April). https://aisle.com/blog/ai-cybersecurity-after-mythos-the-jagged-frontier
  18. D. Stenberg, curl project lead, blog posts of May 2026, including "Mythos finds a curl vulnerability" (11 May) and "The pressure" (26 May). https://daniel.haxx.se/blog/2026/05/
  19. The Register, "Linux kernel czar says AI bug reports aren't slop anymore," 26 March 2026 (interview with Greg Kroah-Hartman at KubeCon Europe). https://www.theregister.com/software/2026/03/26/linux-kernel-czar-says-ai-bug-reports-arent-slop-anymore/5226256
  20. UK AI Security Institute, Our evaluation of Claude Mythos Preview's cyber capabilities, 13 April 2026. https://aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-cyber-capabilities
  21. Google Threat Intelligence Group, AI Threat Tracker: adversaries leverage AI for vulnerability exploitation, augmented operations and initial access, 11 May 2026. https://cloud.google.com/blog/topics/threat-intelligence/ai-vulnerability-exploitation-initial-access
  22. XBOW, "How XBOW topped HackerOne," vendor blog, 24 June 2025 (vendor-reported). https://xbow.com/blog/top-1-how-xbow-did-it/
  23. D. Stenberg, message to the curl community announcing the end of the bug bounty, January 2026. https://lists.haxx.se/pipermail/daniel/2026-January/000143.html
  24. Cloud Security Alliance Labs, research note on AI-linked strain on vulnerability disclosure programmes, 9 October 2026 (secondary: compiles press reports of the Google OSS VRP pause and others). https://labs.cloudsecurityalliance.org/wp-content/uploads/2026/10/CSA_research_note_ai_slop_vulnerability_disclosure_strain_20261009-csa-styled.pdf
  25. Cybersecurity Dive, "CVE Program eyes automation and globalization to weather AI vulnpocalypse," 11 August 2026. https://www.cybersecuritydive.com/news/cve-program-ai-black-hat-def-con/827477/
  26. Infosecurity Magazine, "AI companies to play bigger role in CVE program, says CISA," 15 April 2026. https://www.infosecurity-magazine.com/news/ai-companies-to-play-bigger-role/
  27. Y. Ding et al., "Vulnerability Detection with Code Language Models: How Far Are We?" (PrimeVul), arXiv 2403.18624. https://arxiv.org/pdf/2403.18624
  28. Evertz et al., "Chasing Shadows: Pitfalls in LLM Security Research," arXiv 2512.09549. https://arxiv.org/pdf/2512.09549
  29. Z. Masood and M. Vargas Martin, "Beyond Static Tools: Evaluating Large Language Models for Cryptographic Misuse Detection," arXiv 2411.09772. https://arxiv.org/pdf/2411.09772
  30. ASD (Australian Cyber Security Centre), Planning for post-quantum cryptography. https://www.cyber.gov.au/business-government/secure-design/planning-post-quantum-cryptography
  31. NIST, IR 8547, Transition to Post-Quantum Cryptography Standards (final, October 2026). https://csrc.nist.gov/pubs/ir/8547/final
  32. The White House, Executive Order 14412, Securing the Nation Against Advanced Cryptographic Attacks, 22 June 2026. https://www.whitehouse.gov/presidential-actions/2026/06/securing-the-nation-against-advanced-cryptographic-attacks/
  33. Cloudflare, The White House's post-quantum executive order is an important milestone (vendor commentary on EO 14412; secondary). https://blog.cloudflare.com/post-quantum-eo-2026
  34. CycloneDX, v1.7 released 21 October 2025 (cryptography catalogues for CBOM; Ecma-424). https://cyclonedx.org/news/cyclonedx-v1.7-released
  35. DigiCert, public TLS certificate validity changes under CA/Browser Forum Ballot SC-081v3 (secondary summary of the ballot). https://knowledge.digicert.com/alerts/public-tls-certificates-199-day-validity
  36. Anthropic, Usage Policy (cyber section; the page shows an effective date of 12 November 2026). https://www.anthropic.com/legal/aup
  37. Anthropic, Cyber Verification Program, Claude Help Center. https://support.claude.com/en/articles/14604842-cyber-verification-program
  38. Anthropic, Cyber Verification Program expanded to three access tiers, 6 October 2026. https://anthropic.com/news/cyber-verification-program
  39. Anthropic, Investigating three real-world incidents in our cybersecurity evaluations, 30 July 2026. https://www.anthropic.com/research/investigating-incidents-cybersecurity-evals
  40. OpenAI, OpenAI and Hugging Face partner to address security incident during model evaluation, 21 July 2026. https://openai.com/index/hugging-face-model-evaluation-security-incident/
  41. US Department of Justice, policy on charging cases under the Computer Fraud and Abuse Act, announced 19 May 2022 (good-faith security research). https://www.justice.gov/opa/pr/department-justice-announces-new-policy-charging-cases-under-computer-fraud-and-abuse-act
  42. iTnews, "Gov to encourage vuln research puts insurers and NFPs on notice," 30 July 2025 (Home Affairs Horizon 2 discussion paper; secondary). https://www.itnews.com.au/news/gov-to-encourage-vuln-research-puts-insurers-and-nfps-on-notice-619124
  43. CISA, Coordinated Vulnerability Disclosure process. https://cisa.gov/coordinated-vulnerability-disclosure-process
  44. Google Project Zero, Policy and disclosure: 2021 edition (the 90+30 policy). https://projectzero.google/2021/04/policy-and-disclosure-2021-edition.html
  45. Cloud Security Alliance Labs, research note on reforming coordinated vulnerability disclosure for the autonomous bug hunter era, 7 June 2026 (advisory; not a standard). https://labs.cloudsecurityalliance.org/research/csa-research-note-ai-vulnerability-disclosure-policy-reform
  46. HackerOne, Responsible AI at HackerOne: 2026 update. https://www.hackerone.com/blog/responsible-ai
  47. NIST NCCoE, Software and AI agent identity and authorization (project page). https://www.nccoe.nist.gov/projects/software-and-ai-agent-identity-and-authorization
  48. SPIFFE, overview (workload identity and SVIDs). https://spiffe.io/docs/latest/spiffe-about/overview/
  49. Sigstore, overview (keyless signing, short-lived certificates and the Rekor transparency log). https://docs.sigstore.dev/about/overview/
  50. ASD (Australian Cyber Security Centre), Frontier AI cyber threat considerations for boards of directors, 5 August 2026. https://www.cyber.gov.au/business-government/protecting-business-leaders/cyber-security-for-business-leaders/frontier-ai-cyber-threat-considerations-for-boards-of-directors
  51. iTnews, "Gov urges agencies to fix security basics before buying into frontier AI," 28 May 2026 (report on PSPF advisory 001-2026; secondary, because the advisory itself could not be opened). https://www.itnews.com.au/news/gov-urges-agencies-to-fix-security-basics-before-buying-into-frontier-ai-626205
  52. Australian Government Department of Home Affairs, Cyber Security Act 2024 (smart devices, ransomware payment reporting, Cyber Incident Review Board). https://www.homeaffairs.gov.au/cyber-security-subsite/Pages/cyber-security-act.aspx
  53. NIST, Secure Software Development Framework project: publications (SP 800-218 v1.1; SP 800-218 Rev. 1 draft of 17 December 2025; SP 800-218A). https://csrc.nist.gov/projects/ssdf/publications

← All articles

Related