GPT-5.5-Cyber

Summary

GPT-5.5-Cyber is a specialized variant of [[OpenAI/GPT-5.5|GPT-5.5]] tuned for authorized cybersecurity workflows, released May 7, 2026 in limited preview to a vetted group of cybersecurity teams responsible for securing critical infrastructure.

Overview

GPT-5.5-Cyber is a specialized variant of [[OpenAI/GPT-5.5|GPT-5.5]] tuned for authorized cybersecurity workflows, released May 7, 2026 in limited preview to a vetted group of cybersecurity teams responsible for securing critical infrastructure. It is OpenAI's direct response to [[Anthropic/Claude Mythos Preview|Claude Mythos Preview]] (April 7, 2026), which set the precedent for capability-restricted frontier models gated behind an enterprise trust framework.

GPT-5.5-Cyber is the most permissive model in OpenAI's cybersecurity lineup. It can help defenders write proofs of concept for vulnerabilities they discover, run red-team simulations, hunt for bugs, study malware, and reverse-engineer attacks — but is still hard-blocked from credential theft, malware authoring, and other offensive use cases that fall outside authorized defender workflows.

Specifications

  • Developer: OpenAI
  • Model String: gpt-5.5-cyber (preview)
  • Release Date: May 7, 2026 (limited preview)
  • Type: Cyber-specialized variant of GPT-5.5
  • Access Model: "Trusted Access for Cyber" — vetted enterprise/government partners; account-level controls; mandatory Advanced Account Security from June 1, 2026
  • Distribution: OpenAI API (gated), select managed deployments

Capabilities

Defender-Permissive Tuning: Tuned to be more permissive than the standard GPT-5.5 on security-related tasks — vulnerability research, exploit PoC drafting for organizations' own systems, malware analysis, reverse engineering, controlled red-team simulations, and authorized penetration testing.

Trusted Access Framework: Comes with stronger verification, account-level controls, and individual user attestation; preview access focused on advanced security operations such as authorized red teaming and controlled validation.

Offensive Use Hard-Blocks: Refuses tasks like credential theft, malware authoring, and other clearly offensive cyber actions even for trusted accounts.

Capability Floor (per OpenAI): OpenAI states the preview is not intended to significantly increase cyber capability beyond GPT-5.5; it is primarily meant to remove false-refusal friction in legitimate defender workflows.

Limitations

This is a preview, narrowly scoped to vetted partners. The exact pricing and model card details are not publicly disclosed. GPT-5.5-Cyber inherits the underlying GPT-5.5's hallucination and tool-use failure modes — the safety win is in policy permissiveness, not in step-change capability gain. UK AISI and the U.S. AISI both evaluated the model class and noted that "trusted access" is only as strong as the partner-vetting and account-control regime around it.

The release also intensified scrutiny on how labs handle cyber-capable frontier models. Operational technology (OT) providers reportedly expressed annoyance at being excluded from the parallel Anthropic Mythos rollout, and the broader "cyber-permissive" model category is now squarely on the Trump administration's AI-oversight agenda.

Recent Developments

  • April 7, 2026 — Anthropic's Claude Mythos Preview becomes the first frontier-class cyber-capable model gated behind a restricted partner program (Project Glasswing).
  • May 4, 2026 — Mythos's autonomous-exploit capabilities reportedly become a primary catalyst for new federal AI legislation discussions.
  • May 5, 2026 — Microsoft, Google, and xAI agree to let the U.S. government test their AI models prior to launch.
  • May 7, 2026 — OpenAI announces GPT-5.5-Cyber under the "Trusted Access for Cyber" framework.
  • June 1, 2026 — Advanced Account Security required for all GPT-5.5-Cyber users.

Last Updated

May 9, 2026

→ Back to Models