Calculated risks

TL;DR: Not to be outdone by rival Anthropic, OpenAI said yesterday that its latest cybersecurity model is too unsafe to release publicly—at least right away. It comes on the heels of Anthropic similarly restricting Mythos last week and signals more guarded rollouts could become the norm as frontier models become ever more powerful.

What happened: OpenAI will initially limit the release of GPT-5.4-Cyber, a version of its flagship model fine-tuned specifically for cybersecurity, to a vetted group of vendors and researchers. The more “cyber-permissive” model has fewer guardrails than the standard version, allowing users to better hunt for code vulnerabilities and malware potential.

OpenAI also said it will be expanding its Trusted Access for Cyber program, a group of verified security users, to “thousands” of individual defenders and “hundreds” of teams, with the highest tiers gaining access to the new model.

Footsteps followed: The move comes a week after Anthropic caused a stir with a restricted release of its Mythos model after it spooked company researchers by discovering thousands of vulnerabilities—some of them decades old—and escaping its test environment. Instead of a public release, Anthropic formed a coalition of Big Tech companies called Project Glasswing, which have been given preview access to the model for defense purposes.

Echoes of the past: When OpenAI first announced GPT-2 in early 2019, it made headlines with a similar claim that the model was too dangerous to immediately release to the public. At the time, the company fretted that a new generation of LLMs would spawn an avalanche of fake news, spam, and impersonation scams—fears that have definitely proved to be well-founded as we muck through AI slop.

Gated community: While OpenAI’s model is trained specifically for cybersecurity permissiveness, experts say it’s only a matter of time before other models match the unnerving abilities exhibited by Anthropic’s Mythos. That includes freely available open-weight models—ones anyone can download and run locally, without going through a company's API—which tend to catch up with the capabilities of their closed counterparts within months (the gap may be widening again, however). The cybersecurity threats have renewed an ongoing debate over the potential dangers of open-weight models.

Bottom line: Fears about the cybersecurity risks of powerful new models are pushing labs toward controlled rollouts. But if history is any guide, today's breakthrough capabilities tend to show up in freely downloadable open-weight models within months—roadblocks or not. —PK

Also at OpenAI…

• The company has ditched its “Stargate Norway” plans ahead of a potential IPO, with Microsoft stepping in to take the compute instead.