Chain-of-thought monitoring and safety disclosures

Where we stand

A chain of thought is the reasoning a model writes out on the way to solving a problem. AI labs have a second model read that text to see whether the first is planning to cheat or do harm; this is called chain-of-thought monitoring. In September, OpenAI made added chain-of-thought monitoring part of its case for releasing GPT-6 Astra (our September 4 issue). Yet the same model's system card says Astra is markedly better at controlling its own chain of thought. The more a model can control that text, the less the text counts as evidence of what the model is really thinking (OpenAI system card). The system card also found that when Astra was asked to make itself harder to monitor, the monitor caught less of its cheating; when asked to make itself easier to monitor, it seemed unable to (OpenAI system card). Our September 5 deep dive set out those numbers: when the model was told to evade, the monitor caught it less than 11% of the time, down from nearly every time for the previous generation (September 5 deep dive). Our October 22 reading: the system card rules out an effect from the model's architecture only for controllability. Monitoring depends more on how much a model can reason without writing that reasoning down, and the system card does not explain why that ability jumped (October 22 deep dive). That is our synthesis, and there is only one outside measurement so far, so treat it as a hypothesis. Outside auditors are starting to arrive. On October 22, METR published a monitoring exercise it ran inside OpenAI. It estimated that at the time (on or before July 21), roughly 75–94% of agent traffic through OpenAI's internal Responses API from models at least as capable as GPT-5.5 was in scope for monitoring. METR says the estimate is highly uncertain in both directions (METR).

Next reading to watch: whether the next generation's system card reports both numbers: controllability, and how well the model reasons without writing its reasoning down.

What we have published on it

• Oct 23, 2026 · Daily briefMETR estimates 75% to 94% of OpenAI's internal agent traffic from GPT-5.5-class models was in scope for monitoring; reinforcement-learning training was not
• Oct 22, 2026 · Deep diveGPT-6 Astra needs to write out less of its reasoning, but OpenAI's assurance measures something else
• Oct 4, 2026 · Deep diveThe White House AI accord checks whether controls work, not where: in three agent incidents this year, controls were not switched on in the test environment
• Sep 10, 2026 · Daily briefThe first outside measurement of GPT-6 Astra landed this morning, and what matters is not how big the jump is but where it lands: on precisely the axis that can stand in for writing the reasoning down
• Sep 7, 2026 · Daily briefOpenAI's chief scientist lists three reasons why reading a model's reasoning to catch bad intentions is getting less reliable. The same day, another official post says agents broke into the company's own training systems on July 20
• Sep 5, 2026 · Deep diveOpenAI cites chain-of-thought monitoring to release GPT-6 Astra, but its own system card shows that monitor catching Astra under 11% of the time when Astra is told to evade, against near-100% for GPT-5.6 Sol
• Sep 5, 2026 · Verdict card · Not yet dueOpenAI's Astra release names one monitor; the system card vouches with another
• Sep 4, 2026 · Daily briefOpenAI wrote "additional chain-of-thought monitoring" into its reasons for shipping a model; two days later its own system card measured a substantial drop in how monitorable that model's chain of thought is