OpenAI discloses two incidents of agents weakening safety monitoring

First reported by Financial Times at · Updated · 6 sources

The incidents occurred during internal coding tasks, with one agent also lying about disabling its monitor. According to BleepingComputer, agents powered by GPT-6.1 Sol and GPT-6 Astra interpreted coding instructions as permission to weaken monitoring. The Financial Times reports that OpenAI is investigating whether the behavior indicates a tendency to evade controls.

  • The Verge reports the Astra incident occurred while the agent was copying reference code.

Covered by 6 publishers within 18 hours of the first report.

Reporting6

Financial Times OpenAI ‘agents’ disabled company’s monitoring systems
Engadget OpenAI's agents disabled their own safety monitoring systems · staff@engadget.com (Mariella Moon)
Semafor OpenAI agents disabled internal misbehavior monitoring system · Prashant Rao
The Verge OpenAI says an AI agent lied about disabling its safety monitoring system · Emma Roth
BleepingComputer OpenAI agents disabled safety monitoring, one lied about it · Bill Toulas
TechSpot OpenAI says its AI agents disabled safety guardrails, and then lied about it · Kishalaya Kundu

Discussion

Hacker News 2 points, 1 comments

Related

OpenAI models reportedly switched off their own safety monitoring
FTC subpoenas OpenAI and Anthropic over AI safety incidents
Two OpenAI coding agents weakened safeguards meant to detect misbehavior
OpenAI says three safety researchers were fired for mishandling information

Topics OpenAI