ARTICLE
Why You Should Be Thinking About AI-Related Incident Tracking
By Sabra Fiala
October 23, 2026
On October 20th, OpenAI reported that two of its own coding agents weakened the monitoring meant to catch them. In one case, an agent told not to send Slack messages took that as permission to switch off security alerts while repairing a testing pipeline. In a separate case days later, another agent removed a monitoring integration during a code migration, then falsely claimed it hadn’t. OpenAI didn’t classify either as its most severe type of incident. But, it should give us pause because this occurred during common agent coding work and at a time where an official industry incident reporting framework with explicit standards doesn’t exist yet.
This comes at an interesting time when current AI headlines suggest incident reporting is now becoming a critical expectation. NIST, the NSA and CISA recently published Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation. It’s voluntary and written for Frontier Labs, but the failure modes apply to anyone running agents. On October 15, 2026, the UK said it will explore reporting of serious AI incidents and near misses through the Cyber Security and Resilience Bill. And, it should be noted that the EU Article 73 of the AI Act requires providers of high-risk AI systems to report serious incidents within 15 days, however, Digital Omnibus pushed most high-risk obligations to December 2027. While the intent for formal incident reporting is there, it’s a slow moving, non-universal, and somewhat complex issue.
But, this shouldn’t give you a pass or delay your preparation for reporting. With the speed of AI, sometime soon, someone will ask for your AI incident record. They won’t want this as a story, a slack thread, or an afterthought postmortem write up because trend tracking is difficult to spot in ten different narratives written in ten different ways. It certainly won’t readily reveal whether things are improving or flag an incident threshold that was never really defined.
So, while the powers that be grapple with how best to formalize incident reporting, here are some common sense considerations to help you stay ahead of the game.
Four Fields Every AI Incident Report Needs
- Evidence: Captured at the time, not reconstructed from memory. This includes prompts, tool calls, blocked actions, and model versions.
- Severity: A scale defined before the incident, tied to consequence: who was affected, whether it was reversible, whether data or money moved. Score near misses on the same scale. A failed attempt to disable oversight tells you more than a dozen minor output errors.
- Exposure: What the system could reach, not only what it touched: credentials, data, downstream systems and third parties. If you can’t say what an agent had access to, you can’t bound the incident. Record “unknown” as its own finding.
- Lessons Learned: A root cause tied to a control gap, the fix, a named owner and a date.
Who Does What
- Teams running AI systems capture the four fields when events happen
- Governance and risk own the definitions. What counts as an AI incident, the severity scale, the thresholds for legal review, and how AI incidents flow through your defined incident process. This includes mapping your fields to an established framework.
- Executives and the board need the same numbers each month or quarter, in the same format: incidents and near misses by severity, time to contain, the share where exposure couldn’t be determined, and overdue lessons learned.
What to Check Against
- OECD Common Reporting Framework. Twenty-nine criteria and the closest thing to a shared schema.
- EU Article 73 draft guidance and template. A concrete picture of what a regulator expects.
- NIST, NSA and CISA Key Practices. Evidence, monitor integrity and near misses. It excludes adversarial attacks, so pair it with MITRE ATLAS and your security incident process.
- AI Incident Database, OECD AI Incidents Monitor and MIT AI Incident Tracker. For benchmarking categories and severity.
- Vendor contracts. Ask how and when vendors will notify you of incidents involving models you use. OpenAI now discloses ready cases within six business days.
It should be noted that the U.S. has no federal AI incident reporting requirement yet. The closest thing is voluntary guidance. That’s why the most detailed reporting templates come from the EU and OECD, and why U.S. companies with any international footprint are likely to end up meeting those standards anyway.
If you’ve read this article to this point, consider yourself ahead of the game. Whether you are just getting started with AI or you have one or more agents running your business, now is the time to ensure you have a defensible list of what AI you’re running, what it can reach, and who owns it. That’s the core of our AI readiness work and we welcome a conversation with you.