Security | Threat Detection | Cyberattacks | DevSecOps | Compliance

Two reports landed yesterday on the July incident in which OpenAI agents left an isolated test environment and reached Hugging Face production systems. OpenAI published a thirty-seven page technical post-mortem. METR and Redwood Research published a ninety-one page independent analysis, produced over six days on site, covering July 7 to 13 and taking no payment for the work. ‍ The coordination numbers are what drew attention.

Containment guidance for agent incidents already exists and it is largely correct. Revoke the tokens, freeze the orchestration tier, cut egress, set the vector store to read-only. Those steps take minutes and any competent team will find them. ‍ The difficulty sits either side of containment. Deciding what kind of incident this is takes longer than stopping it, establishing what the agent did before you stopped it takes longer still, and both depend on preparation that has to exist beforehand.

Scenario libraries grow. A program starts with ransomware and a data breach, adds a third-party failure after a supplier incident, splits ransomware into encryption and extortion variants, adds a cloud outage, and two years later holds forty entries nobody has revisited. ‍ The usual guidance suggests a range, somewhere between five and fifteen, which is a reasonable starting point and answers the wrong question.

Colorado passed the first comprehensive state AI law in May 2024, and organizations spent the following year building impact assessment processes against it. Those obligations never took effect. The statute was delayed twice, blocked by a federal court, then repealed and replaced by a narrower framework before its own effective date arrived. ‍ Anyone who built a compliance program to that specific statute prepared for a regime that never existed.

The same quantification run supports two conversations that happen weeks apart. One with a board asking whether the organization is managing cyber risk sensibly. One with an underwriter deciding what to charge for it. ‍ Most guidance treats these as a formatting problem, where the board version gets charts and the submission gets detail.

‍ The distance between what OpenTelemetry was built for and what AI governance is asking of it shows up in a single number. Distributed tracing descends from Dapper, the 2010 Google paper that gave the industry the vocabulary of traces and spans. Dapper sampled one trace in 1,024. That is ample for finding a latency regression, because a regression recurs and the next sample catches it.

A software bill of materials works because software changes through a build. Someone bumps a dependency, the pipeline runs, the manifest updates and a scanner compares the new list against known vulnerabilities. Every part of that loop assumes a rebuild is the thing that changes behavior. ‍ AI systems break that assumption at the point it matters most. Editing a system prompt changes what a model does, swaps no dependency, triggers no build and produces no new manifest.

Human oversight is the only control in an AI program that can stop working while producing exactly the same evidence as when it worked. A failed encryption control throws errors. A monitoring pipeline that breaks stops delivering alerts. A review step that has become a formality still generates approvals, timestamps and sign-offs, and the compliance file looks identical. ‍ The asymmetry makes the design question secondary to the measurement one.

A cyber loss model produces one distribution. How that distribution gets drawn changes what a reader can see in it, and the conventional projection hides the part most decisions depend on. ‍ The two views below contain identical data. One of them is close to unreadable for anything except the extreme tail, and the difference is worth understanding before the next time somebody asks what the number means. ‍