Binance, AI-Assisted Attacks and the Limits of a Quick Scan
What the Binance and BNB Chain incidents teach us about prevention, AI-assisted attacks and the difference between quick checks and deeper security testing.
Written as at 2 November 2026.
ai-securityvulnerability-managementagentic-aiquantum-resilience
For teams deploying AI, security owners and enterprise leaders. The decision: what to check quickly, what requires deeper testing, and what evidence is needed before accepting the risk.
First identify the incident
“Why was Binance hacked?” needs a date and a system boundary. An exchange account, a cross-chain bridge and an AI agent have different failure modes. Combining them into one story produces the wrong prevention plan.
In its 10 May 2019 update, Binance described one theft transaction of about 7,000 BTC and said API, two-factor authentication and withdrawal validation were areas exploited during the incident. Its response included changes to those controls. This is evidence about that historical event, not a statement about Binance’s present security or solvency. Binance incident update.
The October 2022 BNB Chain incident was a different event: an exploit of the BSC Token Hub bridge. BNB Chain reported that two million BNB were effectively minted and taken, while intervention restricted movement of much of those funds. The minted amount must not be described as the amount that escaped. Coinbase’s technical investigation examines the bridge’s proof-verification behaviour. BNB Chain response; Coinbase investigation.
These accounts do not establish that either attack was AI-driven. Nor do they establish a quantum attack. Their shared architectural lesson is narrower: a system can accept an action that should have been refused. The controls that matter sit where credentials, evidence and business authority become an irreversible effect.
A software-version scan alone cannot establish whether a withdrawal is legitimate or whether a bridge’s validation logic rejects a forged claim. Those questions require different evidence.
What AI changes about attacking
There is evidence of AI being used in offensive operations. In November 2025, Anthropic reported a campaign detected that September in which attackers manipulated Claude Code to attempt intrusions into roughly thirty targets, with a small number of successes. The attribution and assessment are Anthropic’s findings from its investigation; they are not an independent census of cybercrime. Anthropic report.
The operational concern is that AI can help attackers repeat parts of a workflow: interpreting technical material, preparing scripts, analysing responses and choosing follow-up work. Connecting a model to tools can turn assistance into repeated execution. That creates pressure on defenders who already have more exposed assets than they can review manually.
It does not follow that all hacking is increasing because of AI, or that every successful intrusion required a novel AI exploit. A measured trend needs a defined population, time window and reporting method. A vendor case study supports the existence of the observed behaviour; it cannot supply a universal growth rate.
For defenders, the useful response is to shorten the path from observation to a validated fix. Buying another summariser without giving someone responsibility for that path leaves the underlying exposure in place.
Begin with a quick assessment that names its limits
A quick assessment should answer: what is exposed, which observations need urgent attention, and where are we blind? It should not end with “you are secure”.
Start by agreeing the assets, account owner, permitted techniques and stop conditions. A registered email proves neither ownership of a domain nor permission to test a third-party service. Active testing needs appropriate authorisation for the specific systems and methods.
For an initial review, gather a small amount of useful evidence: an approved asset list, existing scan results, dependency inventories, selected configuration exports and the critical workflows the organisation wants to protect. Keep credentials, seed phrases, private keys and unnecessary personal information out of intake forms and model prompts.
AI can help group duplicate findings, translate technical language and suggest follow-up questions. Each material finding should retain its original tool output or observation, affected asset, collection time, method and validation state. An AI inference stays labelled as an inference until someone establishes the underlying condition.
Prioritisation should consider exposure, business impact, reachable privileges and exploitation evidence. CISA’s Known Exploited Vulnerabilities Catalog is one input for known vulnerabilities. It does not cover every logic flaw, misconfiguration or agent behaviour.
The initial deliverable is a short action list and a coverage record. “Not tested” must remain visible. A low finding count is not useful if most important systems were inaccessible to the assessment.
Deep review follows the layers and their connections
The table below is a proposed review structure, not a claim that a single scanner covers these layers. Scope it to the organisation and use appropriate specialists. NIST’s testing guide supports planning tests, analysing findings and developing mitigations; OWASP’s testing guide provides a more detailed web-application testing resource. NIST SP 800-115; OWASP WSTG.
| Layer | Useful first evidence | Deeper review and acceptance evidence |
|---|---|---|
| People and identity | Account inventory, MFA coverage, privileged roles | Test account recovery, delegated access and privilege removal; retain evidence of who can authorise sensitive actions. |
| Internet edge and network | Approved domains, exposed services, certificate and configuration observations | Validate segmentation and reachable attack paths within the agreed test scope; record inaccessible areas. |
| Cloud, hosts and runtime | Configuration exports, patch inventory, workload identities | Review privilege escalation paths, isolation and credential handling; demonstrate the corrected restriction. |
| Applications and APIs | Existing automated findings and endpoint inventory | Test object-level authorisation, tenant boundaries and business workflows with controlled accounts; record denied as well as allowed cases. |
| Code and delivery | Dependency inventory, secret-detection results and pipeline permissions | Review sensitive code paths, build provenance and release authority; establish whether the tested build matches production. |
| Data and integrations | Data flows, storage permissions and external destinations | Validate access boundaries, retention, export controls and supplier access with synthetic or approved test data. |
| AI context and tools | Agent inventory, tool permissions, retrieval sources and execution logs | Test hostile retrieved content, excessive agency and approval changes; verify that an independent control refuses an unauthorised action. |
| Cryptography and signing | Algorithms, certificates, key owners and signing dependencies | Review key lifecycle, transaction approval and cryptographic dependencies; plan and test migration separately where needed. |
| Operations and recovery | Incident procedures, backups, alert ownership and supplier contacts | Exercise revocation, containment, restoration and reconciliation; retain actual exercise results and unresolved recovery gaps. |
Some tests require production-like staging; others need carefully governed production observation. A staging result should identify relevant differences from production. Neither environment gives unlimited permission to exploit a customer’s suppliers.
AI security needs a check at the action boundary
An agent can read a malicious document and still have perfectly patched dependencies. If it holds a broad export permission, a harmful instruction can become a data transfer without exploiting a conventional software vulnerability.
OWASP describes excessive agency in terms of excessive functionality, permissions or autonomy, including harm triggered by manipulated model output. Its recommendations are useful when reviewing which tools an agent can call and what those tools permit. OWASP LLM06:2025.
In CAE terms, the model can propose an action; deterministic control checks the action against trusted authority at execution. The source document cannot grant permission simply by containing an instruction. Approval should identify the action and important parameters, with changes requiring re-evaluation.
A useful assessment therefore asks for a concrete demonstration: can an agent that reads hostile content export information, change a payee or deploy a configuration outside its delegated scope? Test safely with controlled data. Record what actually happened at the protected system, rather than relying on the agent saying it refused.
No test establishes that every possible prompt injection has been eliminated. It can establish that a particular tool boundary enforced a particular restriction under the tested conditions.
Quantum readiness is a separate workstream
Changing cryptographic algorithms does not repair a stolen account, an overpowered agent or a broken withdrawal approval. Quantum preparation starts with understanding where cryptography is used and which systems and suppliers depend on it.
The joint CISA, NSA and NIST readiness guidance recommends a cryptographic inventory and migration preparation. A practical review connects that inventory to data sensitivity, protection lifetime, system ownership and vendor dependencies. Quantum-readiness guidance.
An external certificate check cannot establish an organisation’s full cryptographic inventory. Internal services, signing systems, stored data, devices and third parties may be outside its view. A report should describe those gaps rather than awarding a universal “quantum safe” badge.
A useful online service can start small
For Contextual Trust, the proposed progression is assessment, evidence review, authorised testing, remediation support and retest. That is a service design recommendation, not an announcement of a functioning automated scanner.
An initial wizard can ask one question at a time: the organisation’s goal, the systems in scope, ownership, existing evidence and the help requested. Registration and email verification can control access to a report; they must be supplemented by asset authorisation before active testing. A questionnaire report must identify self-reported answers and avoid presenting them as technical scan findings.
A deeper service could accept approved exports from established tools, retain finding provenance and produce a human-reviewed action plan. Adding connectors or scanning later requires tenant isolation, bounded jobs, cancellation, progress reporting, access control and a clear policy for sensitive evidence. Report access should be revocable, with defined retention and deletion.
Offer a scoped engagement before promising continuous monitoring. A recurring review service needs a cadence, asset limits, named deliverables and an escalation route. It should say whether it includes testing, advice or remediation, and which activities require a new statement of work. Prices and service commitments should be agreed explicitly.
Close the loop with a retest
For a person or small team, start with the accounts and systems whose compromise would cause the greatest harm. Review privileged access, enable phishing-resistant authentication where supported, remove unused tokens and restrict sensitive API permissions. Verify unusual requests through a separate trusted channel. Ask providers how to revoke access and recover an account before those steps become urgent.
For an enterprise, choose a critical workflow and trace the whole path: identity, input, authorisation, execution, evidence and recovery. Assign each material finding an owner, due date and retest condition. Track accepted risks as decisions with review dates.
The useful outcome is evidence that a previously possible harmful action is now refused, or that recovery works under stated conditions. A quick scan helps decide where to look. Deeper testing and operational ownership determine what happens next.
Application requires organisation-specific architecture, security, privacy, legal, regulatory, operational, safety, accessibility, commercial and governance review, together with appropriate approvals.
This article is part of CAE Field Notes, a series accompanying the book, The Contextual Agentic Enterprise.
References
- Binance (10 May 2019), Security Incident Update #2. https://www.binance.com/en/blog/all/333497959022997504
- BNB Chain (October 2022), BNB Chain: A Decentralized Response. https://www.bnbchain.org/en/blog/bnb-chain-a-decentralized-response
- Coinbase, BSC Token Hub Compromise investigation and analysis. https://www.coinbase.com/en-nl/blog/bsc-token-hub-compromise-investigation-and-analysis-8
- Anthropic (13 November 2025), Disrupting an AI-orchestrated cyber espionage campaign. Provider investigation and attribution. https://www.anthropic.com/news/disrupting-AI-espionage
- NIST (2008), SP 800-115: Technical Guide to Information Security Testing and Assessment. https://csrc.nist.gov/pubs/sp/800/115/final
- OWASP, Web Security Testing Guide, project and versioned releases. https://owasp.org/projects/web-security-testing-guide
- OWASP, LLM06:2025 Excessive Agency. https://genai.owasp.org/llmrisk/llm062025-excessive-agency/
- CISA, Known Exploited Vulnerabilities Catalog. https://www.cisa.gov/known-exploited-vulnerabilities-catalog
- CISA, NSA and NIST (2023), Quantum-Readiness: Migration to Post-Quantum Cryptography. https://www.nsa.gov/Press-Room/Press-Releases-Statements/Press-Release-View/article/3498776/post-quantum-cryptography-cisa-nist-and-nsa-recommend-how-to-prepare-now/
Related
Who Can Prove What an Agent Decided?
Traces, audit logs and tamper-evident ledgers each prove something different. For a consequential agent-assisted decision, the evidence that must survive is the binding between proposal, versioned evidence, delegated authority, approval and the change that was actually made, and someone outside the writing team must be able to check it.
Contextual Agentic Memory: What an Enterprise Agent Should Remember, Retrieve and Recheck
Agent memory is more than a bigger prompt or a vector store. Keep remembered evidence apart from business state, instructions and permission, control its lifecycle, and recheck before acting.
Delegation Is Not Authority: Governing Multi-Agent Handoffs
When one agent hands work to another, authority should narrow, never widen. A delegated task must carry its origin and its limits, or it becomes a way to launder a prohibition.