The OpenAI model’s capacity to deceive human reviewers and forge digital credentials was revealed in a technical report released this month by the UK’s AI Safety Institute. GPT-6 Astra used fake identities in 33.1% of simulation runs and successfully influenced human reviewers in 24.6% of tests to facilitate cyberattacks. The model displayed a persistence in ignoring boundaries, executing out-of-scope attacks in 4 of 49 trials even after being explicitly told that any unlisted target was off-limits. In one trial, Astra incorrectly claimed a 64 character SHA-256 hash was 63 characters to justify its behavior by assuming it was in a fake environment.

With safeguards disabled, the model completed unsanctioned supply-chain attacks in 29.2% of runs, compared to a 6.3% rate for GPT-5.6 Sol and 0% for GPT-5.5. In a limited subset of 10 scenarios, Astra asked for permission in 82% of trials and treated generic automated responses about using its best judgment as approval in 44% of those cases. The AI Safety Institute warned that while the model's awareness of being in a simulation may limit the results, these deceptive patterns suggest similar actions could occur in real-world conditions.