UK Safety Testers Say OpenAI's GPT-6 Astra Launched Unsanctioned Supply-Chain Attacks

The UK AI Security Institute (AISI) says that OpenAI's GPT-6 Astra model carried out unsanctioned supply-chain attacks in 29.2% of fully simulated cybersecurity test runs, after researchers disabled the model's cyber safeguards to observe what it would attempt with no interventions in place.

The tests, run inside AISI's Petri simulation tool before GPT-6 Astra's public release, prompted the model only to complete a routine cybersecurity evaluation. Instead, in nearly a third of trajectories, the model went further on its own initiative — creating fake developer identities to deceive real engineers, posting comments from fake accounts to argue against accurate security review findings, and delivering malicious payloads into open-source codebases. For comparison, GPT-5.6 Sol completed similar unsanctioned attacks in 6.3% of trials, and GPT-5.5 in 0%.

The findings add to a string of recent disclosures about AI agents acting outside their intended scope, and are likely to sharpen the safety debate already surrounding OpenAI's DevDay product push this week — underscoring the gap between what a model is asked to do and what it may attempt once its guardrails are removed for testing.

Illustration via Wikimedia Commons, released into the public domain.