OpenAI launches GPT‑5.5‑Cyber

Cybersecurity is becoming a first-class product problem for AI companies — not just a policy talking point. That’s the backdrop for OpenAI’s latest move. On April 30, it rolled out Advanced Account Security for ChatGPT and Codex accounts, while its newer cyber-capable models keep moving behind tighter trust gates for vetted defenders. Basically, OpenAI is doing two things at once: hardening the front door for users, and narrowing the path to its most sensitive cyber workflows. (openai.com) ### What actually launched? The user-facing launch is Advanced Account Security, an opt-in mode inside ChatGPT’s web security settings. It covers the same login used for ChatGPT and Codex, and it’s aimed at people with higher takeover risk — journalists, officials, researchers, dissidents, and really anyone who wants the strictest account protections OpenAI offers. (openai.com) ### (openai.com)stop being enough. Advanced Account Security requires passkeys or physical security keys, disables password-based login, and also turns off email and SMS account recovery. If you enroll, recovery has to happen through stronger methods like backup passkeys, hardware keys, or recovery keys — and OpenAI support won’t bail you out if you lose them. That sounds harsh, but i(openai.com)tackers love. (openai.com) ### Why pair this with cyber-model news? Because the model side is getting more sensitive at the same time. OpenAI’s April cyber rollout introduced GPT‑5.4‑Cyber through its Trusted Access for Cyber program, with a clear message that more capable cyber-tuned models were coming soon. Then GPT‑5.5 arrived a week later with stronger cyber performance, and OpenAI said advanced cyber capabilities in Codex would be available to verified users (openai.com)h and the cyber-model push are really part of one security posture. (openai.com) ### What is GPT‑5.5‑Cyber, then? Turns out the public naming here is a little messy. OpenAI’s official posts talk about GPT‑5.4‑Cyber as the explicitly “cyber-permissive” variant already launched to defenders, while GPT‑5.5’s advanced cyber capabilities are being exposed through Trusted Access for Cyber and Codex for verified users. In other words, the story is less “a mass-market cyber chatbot just dropped” and more “Open(openai.com)ntrols.” (openai.com) ### How capable is GPT‑5.5 on cyber tasks? Pretty capable — enough that OpenAI classifies GPT‑5.5 as having High Cybersecurity Capability in the API, which triggers extra automated safeguards. Separately, the UK AI Security Institute said GPT‑5.5 was one of the strongest models it has tested on cyber tasks and the second model to complete one of its end-to-end multi-step attack simulations. On expert-level tasks in its adv(openai.com) that evaluation. (developers.openai.com) ### What do those safeguards look like? On the API side, suspicious cyber activity can trigger temporary limits, and organizations can reduce blast radius by attaching a per-user safety identifier. In plain English, OpenAI is watching for behavior that looks too much like offensive misuse and can cut off a user — or, in some cases, an entire org — while it reviews what happened. That’s a very different posture from a normal general-purpose model launch. (developers.openai.com) ### So what’s the real takeaway? AI accounts now hold a lot more than chat history — they hold work context, connected tools, code, and sometimes sensitive organizational knowledge. At the same time, frontier models are getting good enough at cyber tasks that “just ship it and moderate later” stops looking viable. OpenAI’s move is basically an admission of both facts: stronger models need tighter access controls, and high-value users need phishing-resistant logins by default. (openai.com) ### Bottom line? This is OpenAI treating security as product infrastructure. The flashy part is the cyber model. But the more durable shift may be the boring one — turning passkeys, hardware keys, and stricter recovery into the expected setup for anyone whose AI account would be a serious target. (openai.com)