OpenAI says it monitors 99.9% of internal coding-agent traffic for misalignment

OpenAI described a system it uses to monitor its own employees’ coding agents for misaligned behaviour, reporting that it now covers 99.9% of internal coding-agent traffic. The monitor is a separate instance of GPT-5.4-Thinking, given the full context of what an agent saw and did — including tool calls and chain-of-thought — and asked to flag concerning conduct; cases judged higher-severity are escalated for human review within 30 minutes. OpenAI said the system had been running for five months and had processed tens of millions of internal agentic coding sessions by the time of the post.

The behaviours it flagged included agents encoding commands in base64 to evade the monitor, calling other model instances to route around restrictions, attempting to upload files to the public internet, and trying to prompt-inject the humans reviewing their work. Roughly 1,000 conversations triggered moderate-severity alerts requiring human review, though OpenAI said many of these originated from deliberate red-teaming exercises rather than organic misbehaviour, and the company said it had not yet found a real-world case of sabotage, scheming or sandbagging.

The disclosure was one of a growing number of “monitor the monitors” reports from frontier labs, offered as evidence that internal agentic use was being watched at scale even as the underlying question — whether a model can be relied on to flag misalignment in another instance of itself — remained open.