OpenAI expands GPT-5.5 for cybersecurity

OpenAI is expanding access to GPT-5.5 for verified cybersecurity teams and introducing GPT-5.5-Cyber, a limited-preview model for more specialized defensive tasks.

The difference is not just the model. It is also who can use it, which systems they can use it with, and what controls are in place.

Three levels of access

The company has created Trusted Access for Cyber (TAC), a system based on the identity and trust level of each user or organization. Its goal is to reduce unnecessary blocks on defensive work without enabling uses that could cause real harm.

OpenAI has defined three levels:

• Standard GPT-5.5: maintains the usual protections and is intended for general use, development, and knowledge tasks.
• GPT-5.5 with TAC: provides more useful responses for verified defenders working in authorized environments. It is designed for reviewing code, analyzing malware, studying vulnerabilities, creating detections, and validating patches.
• GPT-5.5-Cyber: enables more flexible behavior for specialized tasks, such as penetration testing and controlled red team exercises. For now, it is available only in limited preview.

In all cases, requests related to credential theft, persistence, malware deployment, stealth, or the exploitation of third-party systems remain blocked.

What changes for security teams

For most defenders, OpenAI recommends starting with GPT-5.5 and TAC. The model can help understand unfamiliar code, identify the cause of a vulnerability, review a fix, prioritize the severity of a flaw, and prepare a safe reproduction in an isolated environment.

This can shorten tasks that normally take hours. For example, an analyst could provide a project from their own software, ask the model to trace an attack path, and use the results to check whether a patch actually fixes the problem.

GPT-5.5-Cyber is reserved for higher-risk situations. A team conducting an authorized penetration test may need to check whether a vulnerability is exploitable, not just receive a theoretical explanation. In those cases, refusals from the general model can interfere with the work, and the new model is intended to provide more room under stricter controls.

OpenAI clarifies that this first version is not designed to outperform GPT-5.5 in every cybersecurity evaluation. Its main difference is that it is more permissive on security-related tasks, not that it necessarily has greater capability in every test.

More verification for greater capability

Expanded access requires additional measures. Individual users who employ the more permissive models will have to enable Advanced Account Security, a phishing-resistant account protection, starting June 1, 2026.

Companies may instead demonstrate that they use phishing-resistant authentication within their corporate login system. OpenAI also mentions controls for authorized use, monitoring for potential abuse, and limits tied to organizations and target systems.

The idea is simple: the more sensitive the task, the more trust there must be in who is carrying it out and in the environment where it takes place.

A network of defensive tools

OpenAI is testing these capabilities with companies working in networking, detection, vulnerability analysis, and software supply chain security. Planned uses include:

• Reviewing firewall rules and configurations while preparing a patch.
• Analyzing alerts and telemetry from EDR and SIEM tools, which monitor devices and security events.
• Detecting vulnerable dependencies or compromised packages before they reach production.
• Generating remediation instructions and validating code changes.

The company is also expanding Codex Security, which creates a threat model specific to each codebase, analyzes possible attack paths, and proposes patches for a person to review. In addition, some open-source project maintainers may receive conditional access and credits to reduce their maintenance burden.

During initial testing, GPT-5.5-Cyber was used to automate red team exercises and validate high-severity vulnerabilities, according to OpenAI. The company says it will publish more details in a later technical analysis as part of responsible disclosures.

For you, the change will depend on your relationship with cybersecurity. If you run a company, these tools can speed up code review, incident investigation, and patch deployment. If you are a developer, they can help you find flaws before release. But access to the most sensitive features will not be open by default: it will depend on identity, authorization, and the controls in the environment.

OpenAI is testing a gradual access model for cybersecurity: general capabilities for most users, more flexible functions for verified defenders, and specialized models for higher-risk cases. The key question now is whether oversight and verification can scale at the same pace as model capability.