OpenAI unveiled GPT-5.5-Cyber, a new AI model focused on cybersecurity, less than a month after rolling out GPT-5.4-Cyber in response to Anthropic's "Claude Misos."
On the 7th (local time), OpenAI released GPT-5.5-Cyber in a limited preview aimed at strengthening cyber defense capabilities.
The company is offering the model to security teams that protect critical infrastructure and to vetted defensive organizations. It is designed to support advanced tasks such as vulnerability analysis, penetration testing and malware analysis.
OpenAI noted that its latest general model, GPT-5.5, released two weeks earlier, already delivers significant cybersecurity capabilities. Through a program called Trusted Access for Cybersecurity (TAC) built on GPT-5.5, the company provides additional, specialized features to verified defensive organizations.
OpenAI emphasized that, for most security teams, GPT-5.5 paired with TAC is the most appropriate starting point. That combination can handle the bulk of defensive work: secure code review, vulnerability prioritization, malware interpretation, detection engineering and patch validation.
TAC operates on identity verification and trust-based attestation. For approved defensive organizations, it relaxes model restrictions for lawful defensive work—vulnerability detection and classification, malware analysis, binary reverse engineering, detection-rule creation and patch verification.
GPT-5.5-Cyber is designed to go further, enabling higher-level security operations.
Access controls and safeguards are central. OpenAI stressed that as cyber AI models become more capable, the risk of misuse grows, making it critical to verify who is using them and for what purpose.
OpenAI said the early preview of GPT-5.5-Cyber does not substantially raise overall performance compared with GPT-5.5. Instead, it has been trained to deliver more permissive responses for security-related tasks. In practice, that lets approved research environments carry out security experiments and exploit verification that the general model might decline as too sensitive.
At the same time, OpenAI is working to strengthen the broader cybersecurity ecosystem.
It proposed a "Security Flywheel" that links vulnerability discovery to patch deployment, detection and response, and network blocking. Under this model, when a researcher finds a vulnerability, supply-chain security tools can block risky code, EDR and SIEM vendors can detect traces of real attacks, and network security firms can push WAF rules and edge-blocking policies to limit damage.
OpenAI said it will also expand support for open-source maintenance. Via Codex Security, the company will help projects generate threat models, map attack paths, validate vulnerabilities and propose patches. It plans to give maintainers of critical open-source projects access to Codex Security and API credits through Codex for Open Source.
OpenAI also released a Codex Security plugin for app and CLI environments. Developers can use it to handle threat modeling, vulnerability discovery and validation, attack-path analysis and patch-code generation within a single workflow.
OpenAI said that in alpha testing GPT-5.5-Cyber is already being used to run automated red-team tests of critical systems and to validate high-risk vulnerabilities. The company added that it will continue developing even more powerful cyber-specialized models.
By Chan Park cpark@aitimes.com
