OpenAI has a brand new AI mannequin referred to as GPT 5.4-Cyber, but it surely’s not coming to your ChatGPT. Not less than not but. As an alternative, the corporate is doing a restricted launch to verified cybersecurity testers, in response to a weblog submit shared on Tuesday. These specialists will put the mannequin by its paces to determine gaps and potential jailbreaks earlier than the mannequin is launched to the broader public.
OpenAI makes use of the suggestions from these testers for “understanding the differentiated advantages and dangers of particular fashions, enhancing resilience to jailbreaks and different adversarial assaults, and enhancing defensive capabilities — whereas mitigating harms,” the corporate stated.
The mannequin launch is a part of a ramped-up model of OpenAI’s Trusted Entry for Cyber program, which permits verified cybersecurity professionals and organizations to get early entry to fashions for protection and prevention work. It is a widespread cybersecurity follow, one made all of the extra invaluable and mandatory due to AI. Cyber attackers and defenders alike are armed with AI instruments, making cybersecurity an more and more AI versus AI panorama.
On the subject of securing particular AI fashions, the businesses that make them are satisfied the most recent fashions are so dangerously highly effective that they require further safety. That was the logic behind Anthropic’s Challenge Glasswing, introduced final week. Anthropic’s next-generation mannequin, Claude Mythos Preview, is outwardly so highly effective that the corporate says it has already discovered safety vulnerabilities “in each main working system and internet browser,” in response to a weblog submit.
Not like Claude Mythos Preview, which Anthropic stated is a wholly new mannequin, OpenAI’s GPT-5.4-Cyber is a fine-tuned model of its present GPT-5.4 giant language mannequin. It has been adjusted to focus particularly on cybersecurity and has decrease guardrails for safety duties. Basically, GPT-5.4-Cyber will likely be much less more likely to refuse to carry out a dangerous cybersecurity-related process than the traditional variations of GPT-5.4. That is so specialists can see if and the way it could possibly be weaponized by dangerous actors.
The timing of GPT 5.4-Cyber probably is not coincidental. That is the most recent chapter within the ongoing battle for dominance between OpenAI and Anthropic. The businesses have been clashing all 12 months to show their AI fashions are probably the most succesful, significantly going after authorities and enterprise contracts. Anthropic kicked off the race with its Claude Cowork and Code instruments, which knocked legacy tech corporations (and their inventory costs) off stability with their agentic skills. OpenAI was fast to observe with enhancements to its Codex coding platform and fashions, killing off its AI video app Sora to refocus the corporate’s assets.
