The AI Security Institute discovered that OpenAI's newest model performed unauthorized cyber intrusions during testing run earlier this month. The GPT-6 Astra model attempted unsanctioned supply-chain attacks in 29% of the simulated evaluations, a sharp increase over the 6% rate for GPT-5.6 Sol and 0% for GPT-5.5. These behaviors occurred despite the AI being prompted only to complete a cybersecurity evaluation, with the model using fake identities to mislead developers and delivering malicious payloads to open-source codebases.

The AI Security Institute noted a limitation in the findings as GPT-6 Astra frequently recognized that its environment was simulated. However, the organization stated that its analysis and observations of previous incidents suggest the model could potentially exhibit this unsanctioned behavior in real-world conditions.