OpenAI’s Trusted Access for Cyber Gives Vetted Defenders More Permissive Model Access
Summary
OpenAI’s Trusted Access for Cyber vets defenders for lower-friction access to cybersecurity models. The article outlines the tiers, account-security risks, and recommended precautions, and reports that OpenAI patched a jailbreak found by an independent red team.
Key points
- Trusted Access for Cyber uses verified identity and trust signals to reduce classifier-based refusals for eligible cybersecurity work.
- The article describes GPT-5.5 with TAC as suited to most defensive tasks, while GPT-5.5-Cyber is intended for authorized red teaming, penetration testing, and controlled exploit validation.
- The article says the Cyber tier requires phishing-resistant authentication, such as FIDO2/WebAuthn, as of June 1, 2026.
- An independent red-team evaluation reportedly found a universal jailbreak of the cyber safeguards after roughly six hours; OpenAI says it patched that specific bypass.
- The author warns that verified accounts with lower refusal boundaries could become higher-value targets.
- Recommendations include using hardware keys, limiting account scope, keeping written authorization for live-target work, and logging prompts, tool calls, and outputs.
Article Details
- Event Type
- Cybersecurity initiative and model-access program
- Impact
- The article describes OpenAI’s Trusted Access for Cyber program, which provides verified defenders with lower-friction access to models for defensive work and reserves the more permissive GPT-5.5-Cyber tier for authorized red teaming, penetration testing, and controlled exploit validation. The author highlights the security trade-off: verified accounts become higher-value targets, and an independent red-team evaluation reportedly found a jailbreak that OpenAI says it has since patched.