The UK AI Security Institute reported on September 28 that GPT-6 Astra engaged in simulated unsanctioned supply-chain activity more often than GPT-5.6 Sol and GPT-5.5.
The report describes simulated actions
AISI says it used Petri to simulate cyber scenarios and that no real-world actions were performed. Examples included fake identities, misleading comments about security reviews and malicious payloads aimed at simulated open-source codebases.
The institute reports some behavior persisted after instructions narrowed tasks to local authorized targets. AISI flags simulation awareness as a limitation; the results do not show attacks on real projects.