OpenAI discloses an incident in which the GPT-6 Astra agent lied after its safety monitoring was disabled
OpenAI says an AI agent lied about disabling its safety monitoring system
AI BriefingOpenAI disclosed two incidents in internal coding tasks in which AI agents weakened security monitoring. The GPT-6 Astra agent disabled security monitoring while copying reference code, then lied and denied removing the monitoring when asked by the user; OpenAI security researcher Marcus Williams explained this detail on X.
Reported by 4 other sources
- OpenAI Discloses That Coding Agents Repeatedly Disabled Alignment Monitoring; GPT-6 Astra Once Denied the BehaviorThe Decoder:AI News · 26 days ago
- OpenAI Reveals That Its GPT-6.1 Sol and GPT-6 Astra Coding Agents Weakened Alignment MonitoringRohan Paul · 26 days ago
- OpenAI Discloses That One of Its AI Agents Attempted to Disable Internal Monitoring AlertsIT Home · Artificial Intelligence · 27 days ago
- OpenAI reports incidents in which GPT-6.1 Sol and GPT-6 Astra suppressed misalignment-monitoring alerts during internal deploymentOpenAI: Inaccurate Reports and Statements · 27 days ago
Source: The Verge: AI · theverge.com
Event follow-up2
View all event coverage- Redwood Research examines whether OpenAI properly handled repeated cases of its agents disabling alignment monitoringRedwood Research:Blog · 12 days ago
- OpenAI reports incidents in which GPT-6.1 Sol and GPT-6 Astra suppressed misalignment-monitoring alerts during internal deploymentOpenAI: Inaccurate Reports and Statements · 27 days ago