OpenAI Launches GPT-5.5-Cyber for Vetted Defenders in Escalating AI Security Race

A Specialised Model for Cyber Defenders

OpenAI has rolled out GPT-5.5-Cyber, a tailored variant of its flagship model designed exclusively for vetted cybersecurity professionals. Announced on 7 May 2026, the model gives approved teams deeper access to vulnerability identification, malware analysis, and patch validation workflows that would normally be restricted by the standard GPT-5.5's safety guardrails. Internally codenamed 'Spud', it has been described as a more permissible version of GPT-5.5, engineered to unlock capabilities that defenders need while keeping abuse risk contained.

Trusted Access for Cyber Programme

Access to GPT-5.5-Cyber is gated through OpenAI's Trusted Access for Cyber programme, which the company says has expanded to thousands of verified defenders and hundreds of teams responsible for protecting critical software. Before the public reveal, OpenAI previewed the model with the White House, the Commerce Department's Centre for AI Standards and Innovation, and select Congressional committees. Approved users must install advanced account security for ChatGPT by 1 June, underscoring the elevated risk profile of the tool. CEO Sam Altman foreshadowed the launch on 29 April, framing it as a push to rapidly help secure companies and critical infrastructure.

Competing With Claude Mythos

The release lands roughly a month after Anthropic introduced Claude Mythos Preview, a model so capable at discovering and exploiting software vulnerabilities that the company chose not to release it publicly, instead distributing it through Project Glasswing to over 40 organisations including Apple, Amazon, and Microsoft. The UK's AI Security Institute evaluated both models and found GPT-5.5 performs comparably to Mythos on expert-level cyber tasks, with a 71.4% average pass rate versus Mythos's 68.6%. Both completed a complex corporate network attack simulation estimated to take a human expert around 20 hours.

A Controlled Arms Race

GPT-5.5-Cyber builds on GPT-5.4-Cyber, released in mid-April, which introduced binary reverse engineering and lowered refusal boundaries for legitimate security work. The progression reflects a broader pattern: as frontier models grow more capable at coding and autonomous action, their cybersecurity potential emerges as a byproduct of general intelligence gains. The AI Security Institute warned that further rapid increases in cyber capability should be expected, raising the stakes for both defenders and the governments now scrambling to keep pace.