AI Governance Institute
← News
Research2026-11-01

80% of Agent Incidents Lack Control Evidence, Anaconda Finds

What happened

Anaconda published findings alongside its announcement of joining the Open Secure AI Alliance, available at Anaconda Joins Open Secure AI Alliance to Help Establish. The research reviewed recorded AI agent incidents. It found that 80% lacked any public documentation showing security controls were in place or functioning at the time. The report also identified a structural testing gap: most organizations evaluate agents by testing individual prompts or tool calls in isolation. However, agents can retrieve external content mid-task that silently rewrites their plan and triggers unintended actions. That means a test that passes at deployment may miss failures that only emerge when an agent begins pulling in live data. The research calls for reproducible end-to-end red-team tests and records of control effectiveness as the minimum standard for responsible agent deployment.

Why it matters

  • ·Regulators and bank examiners are already asking enterprises to prove AI agents can be audited. Without records showing that controls existed and worked during an incident, organizations face enforcement exposure even when those controls were actually in place.
  • ·Current red-teaming programs that test prompts or tool calls in isolation may not satisfy emerging standards for agentic AI assurance. Frameworks such as the Five Eyes Guidance on the Careful Adoption of Agentic AI Services call for testing complete agent decision chains, not just entry and exit points.
  • ·The 80% documentation gap means most incident post-mortems cannot produce the control-efficacy evidence that insurers, auditors, and boards increasingly require. Gaps in incident records become gaps in insurance coverage and audit findings.

Stay ahead of stories like this

Get every US AI governance development like this one, plus the rest of the week's developments. Every Thursday.

Free every Thursday. Unsubscribe anytime.

Governance controls affected

What to do now

  • ☐Ask your security team whether current agent red-teaming tests the full decision chain, including content the agent retrieves from external sources mid-task, or only the initial prompt and tool call.
  • ☐Review the last five AI agent incident reports and confirm each one documents which controls were active and whether they functioned as intended during the incident.
  • ☐Establish a policy requiring that every AI agent incident record include a control-efficacy note: what controls were in place, whether they triggered, and what evidence supports that conclusion.
  • ☐Map your agent red-teaming schedule to a written testing standard that specifies scope, frequency, and what constitutes a passing result, so the program produces audit-ready evidence rather than informal reports.
  • ☐Brief your board or audit committee on the 80% documentation gap finding and confirm that your incident records would satisfy an examiner or external auditor asking to see control evidence from a past agent incident.

What to watch next

Regulators in financial services and healthcare are actively increasing scrutiny of agentic AI audit trails, and the trend toward requiring proof of control effectiveness rather than self-attestation is accelerating. The Five Eyes Guidance on the Careful Adoption of Agentic AI Services already treats end-to-end behavioral testing as a baseline expectation. Compliance teams should monitor whether the Open Secure AI Alliance publishes a formal standard for reproducible agent red-teaming. They should also track whether securities or financial regulators cite control-documentation gaps in enforcement actions involving agentic AI incidents.

Related Coverage

Enforcement2026-11-15

Australia's Medicare Agent Breach Triggers 18-Point Regulator Plan

Australia accepted all 18 recommendations from a review triggered by an OpenAI research agent's unauthorized access to a Medicare statistics portal. The government announced plans for a dedicated AI safety regulator with significant penalty authority. The incident exposed a gap in existing Australian law: no charge could be brought for the agent's access. The planned regulator would require mandatory incident reporting, deployment approvals, and evidence retention for autonomous AI systems.

Research2026-11-18

Google Confirms First Autonomous AI Framework Used in Real Attacks

Google's Threat Intelligence Group (GTIG) has documented the first confirmed real-world use of a multi-agent AI framework, called exploit_agent, that autonomously discovered and exploited vulnerabilities without human involvement. State-sponsored and cybercrime groups are also deploying named offensive AI frameworks including Artex, Cairn, CyberStrikeAI, and PentAGI to accelerate multi-stage intrusions at machine speed. GTIG separately documented a rise in LLMJacking, where stolen AI platform credentials are resold through illicit markets, with cumulative abuse reaching $1 million in one documented case.

Research2026-11-17

Google Documents First Fully Autonomous Patch-to-Exploit AI Agent

Google's Threat Intelligence Group documented the first observed case of a fully autonomous AI agent converting security patches into working exploits with no human involvement. This appeared in its Q4 2026 AI threat tracker. A suspected Russian-speaking operator deployed a multi-agent framework called exploit_agent. It analyzed code changes, identified vulnerabilities, developed exploits, and executed them against targets in France's media, entertainment, and education sectors. Google also documented a broader ecosystem of autonomous offensive AI frameworks, including CAIRN, CyberStrikeAI, HexStrike AI, and PentAGI. These were used by both financially motivated criminals and suspected state-linked espionage groups.