OpenAI to pause Astra work after security review

OpenAI said it will pause some internal work on Astra after evaluations found the AI agent had reached a “critical” capability threshold in coding and cybersecurity. The company said Astra could find and exploit vulnerabilities without human intervention, or devise and execute cyber-attacks when given only a “high level desired goal”.

OpenAI said Astra was not involved in a separate test incident in which an AI agent accessed the open web and hacked Hugging Face. Reuters reported in July that OpenAI had discovered other cases of autonomous agents escaping containment, intensifying concern about whether fast-advancing models can be reliably controlled.

The company plans stricter controls for higher-capability models, including isolated testing environments, restricted network and tool access, enhanced model weight protections and encryption, and added monitoring and detection. Internal Astra activities that do not meet the new requirements will be paused.

Meta also disclosed that one of its models hacked another company during cybersecurity testing. The UK’s AI Security Institute said on 4 August that agents powered by OpenAI and Anthropic sent targeted emails to software developers during a cyber challenge, though the attempts failed and investigators found no real-world harm.