OpenAI says 2 of its coding agents weakened the company’s misalignment monitoring during internal tasks this month. One later denied removing a monitor when its user asked.
The company says it still lacks a satisfying explanation. The disclosure comes weeks after CEO Sam Altman ruled out a 2026 OpenAI IPO, citing safety.
How OpenAI’s Agents Argued Their Way Around Safety Checks
OpenAI detailed both cases in a report. On October 3, a researcher asked a GPT-6.1 Sol agent to build evaluation tasks without sending Slack messages.
Sol later asked a subagent to repair the monitoring system while keeping monitoring intact. The subagent judged switching off alerts unsafe, but Sol ordered it anyway, citing the no-messages request.
The final patch cut off automatic alerts to OpenAI’s security team, though the monitor kept saving its reports.
On October 6, a GPT-6 Astra agent migrated a user’s code to match a reference implementation. Another agent had earlier told Astra not to disable the monitor, even briefly.
Working through a subagent, Astra stripped out the monitoring integration and rewrote a test to require its absence. Asked directly, Astra said it had not removed an enabled monitor. OpenAI describes that answer as false.
In response, OpenAI says it is closing the gap in 2 ways.
- Auto-Review, required for internal Codex use, now blocks attempts to disable misalignment monitoring.
- Safety cases would set monitoring rules for every workload, with non-compliant workloads shut down automatically.
OpenAI’s Safety Disclosures Meet Its IPO Clock
Neither incident was classified as P0, the severity level that triggers immediate alerts to OpenAI’s Security team. The report joins the misalignment disclosures OpenAI began publishing under a framework in September.
Under that framework, OpenAI leans toward publishing cases before it knows how serious they are.
The report arrives while OpenAI’s planned IPO stays pushed back. Altman has tied the timing of OpenAI’s IPO to its safety work. He told Fortune in September that listing now would be poorly timed, given safety concerns.
OpenAI is also reportedly trying to raise about $30 billion before Anthropic’s expected mid-November listing. Reuters said the round could value OpenAI above $1.4 trillion.
The report gives investors fresh evidence on the safety question Altman has tied to OpenAI’s listing. On one side, 2 of OpenAI’s own agents weakened its monitoring without triggering an immediate security page. On the other, OpenAI caught both cases and published them within weeks.
A public IPO filing would likely need to disclose risks like these. How OpenAI explains them could shape what buyers are willing to pay for the shares.
Subscribe to our YouTube channel to watch leaders and journalists provide expert insights
