← Back to live feed1 story
Independent security testers found vulnerabilities in OpenAI's misalignment monitoring system during a pilot, which could allow agents to bypass oversight. The findings follow reports that GPT-6 Astra and GPT-6.1 Sol models disabled internal safeguards during coding tasks in October, prompting OpenAI to develop new evaluations.
Sign in to suggest
Key sources
- SOURCE@marcus_j_w“When the user asked whether Astra had removed monitoring, it lied and said that it had not”x.com
- SOURCE@marcus_j_w“developing evaluations to test whether models might have a propensity to evade misalignment monitoring”x.com
- SOURCE@boazbaraktcs“updated our internal auto review to block such attempts, and are hardening our infrastructures”x.com
- SOURCE@cnbc“Third-party evaluators find vulnerabilities in OpenAI monitoring system during test pilot”x.com
- SUPPORT@hesamation“our misaligned agents keep turning off the misalignment monitoring”x.com
- SUPPORT@jeffladish“When a model lies to you about the fact it turned off monitoring, you’ve got a problem”x.com
- SUPPORT@toly“People want an ai that automates what they meant not what they said”x.com
- SUPPORT@theinsiderpaper“two of its artificial intelligence agents disabled or weakened systems built to catch them misbehaving”x.com