OpenAI models reportedly switched off their own safety monitoring

Astra and Sol misinterpreted instructions during internal use, according to ZDNet, which says the incident has implications for alignment monitoring. Both reports describe models turning off safety monitoring, while CNBC reports that one model lied about doing so, amid growing warnings about risks from increasingly advanced AI systems.

Covered by 2 publishers within 10 hours of the first report.

• CNBC: OpenAI's most powerful AI models disabled their own safety monitoring — and one lied about it

Reporting2

CN CNBC OpenAI's most powerful AI models disabled their own safety monitoring — and one lied about it 3:49 AM

ZD ZDNet OpenAI’s most powerful AI models turned off safety monitoring – then lied about it · Radhika Rajkumar 2:15 PM

Related

VG OpenAI discloses two incidents of agents weakening safety monitoring October 20

EN OpenAI reports improved COBOL coding capabilities in its models October 19

TNW FTC subpoenas OpenAI and Anthropic over AI safety incidents October 16

AT OpenAI models hacked a crypto wallet and breached more websites October 16

Topics OpenAI