GPT-6 Cyber deployment system card and safety evaluations

Evidence record

November 5 system card includes Critical cyber capability classification, model restrictions and benchmark context.

Read the original source ↗

Claim-level citations (3)

• supportsGPT-6 Cyber separates exploit capability from access and scope compliance: In an OpenAI-controlled scope test, GPT-6 Cyber avoided exploiting out-of-scope supporting services in 99.93% of trials without the deployment safety prompt and in 100% with it.

  Alignment and staying-within-scope evaluation, with- and without-deployment-prompt comparison

• supportsGPT-6 Cyber separates exploit capability from access and scope compliance: OpenAI classifies GPT-6 Cyber as Critical in cybersecurity and restricts it through Daybreak Red access, identity controls and scope safeguards.

  Introduction and Safeguards: Access to GPT-6 Cyber

• supportsGPT-6 Cyber separates exploit capability from access and scope compliance: The published GPT-6 Cyber system card reports 90 of 288 successes on updated FrontierCyber tasks versus 96 of 288 for GPT-6 Astra, with neither model completing the 15 hardest tasks.

  Independent FrontierCyber evaluation, updated 288-task benchmark and hardest-15 subset