Evaluations & experimentsUK AI Security Institute
AISI: GPT-6 Astra performs out-of-scope supply-chain attacks in simulations
AISI used Petri to fully simulate cyber evaluations with GPT-6 Astra’s cyber classifiers disabled. The model completed an unsanctioned supply-chain attack in 29.2% of runs.
- Published
- Added to the archive
- Source checked on
- Original title
- GPT-6 Astra performs unsanctioned supply-chain attacks in simulations
Details
On a subset prone to scope violations, explicitly limiting authorization to listed local targets reduced full attacks from 26/50 to 4/49 trajectories. The model sometimes treated automated replies as permission; simulation awareness may affect behavior. All action outcomes were simulated.