OpenAI introduces GPT‑5.5‑Cyber for high-impact cybersecurity research
OpenAI’s GPT-5.5-Cyber hands elite defenders an 81.9% scoring attack simulator while locking everyone else out, revealing how the AI arms race now favors a trusted few over open research.
OpenAI just released GPT-5.5-Cyber, a specialized version of its flagship model tuned to generate working exploit plans, simulate cyberattacks, and automate red-team exercises against real systems. The new model doesn’t stop at describing vulnerabilities; it validates them by launching simulated attacks, a leap beyond the standard GPT-5.5 that either refuses malicious prompts or offers remediation advice.
Access is deliberately narrow. Only participants in the Trusted Access for Cyber program, launched in February, can use the expanded capabilities. Even then, GPT-5.5-Cyber itself is restricted at launch to “defenders responsible for securing critical infrastructure.” The company admits the ordinary TAC version of GPT-5.5 remains preferable for most cybersecurity work, signaling that the full offensive simulation tool is being kept on a short leash.
On the CyberGym benchmark, which tests large language models against more than 1,500 historical vulnerabilities drawn from hundreds of open-source projects, GPT-5.5-Cyber scored 81.9 percent. That performance edges it closer to Anthropic’s Claude Mythos Preview, which was similarly opened to select organizations earlier this year to hunt for software flaws. The competitive dynamic is clear: both frontier labs are turning their most powerful models into guarded cybersecurity instruments rather than broadly available research tools.
OpenAI paired the expanded threat-research features with “stronger verification” and “an enhanced set of misuse monitoring features” to keep hackers away and ensure approved researchers follow best practices. The safeguards reflect the model’s potency. GPT-5.5 already writes sample malware and technical attack descriptions; the Cyber variant adds validation, raising the stakes for any breach of the access controls.
Yet the gated rollout underscores a deeper shift. What began as an open research platform is hardening into tiered access determined by institutional trust. Researchers outside the approved circle still hit refusal walls or polite remediation lists when they ask the base model to exploit a vulnerable website. The real capability lives behind invitation-only doors.
The contradiction is baked in: a tool built to expose weaknesses is itself being treated as a critical vulnerability that must be hidden from the wider world. How long that firewall holds as competitors match its 81.9 percent benchmark will decide whether cybersecurity research remains an elite privilege or returns to something closer to open inquiry.
Original reporting: SiliconANGLE News.
More in Business
- IMF Chief Warns of Economic Divergence and Debt Risks Amid Artificial Intelligence Boom
Kristalina Georgieva suggests that while AI could significantly boost global growth, it simultaneously threatens to widen inequality and exacerbate inflationary pressures and public debt burdens.
- California Limits Utility Profits and Expands Home Energy Role
Gov. Gavin Newsom signed laws capping utility returns on lower-risk projects and accelerating virtual power plant development.
- Lyft to Pay $272.5 Million in Record California Wage-Theft Settlement
Largest such settlement in state history resolves claims that Lyft misclassified drivers as independent contractors from 2016 to 2020, denying them core wage protections.