Frontier AI tools being gated
Leading AI labs are holding back some powerful cybersecurity-focused models instead of releasing them broadly, citing safety concerns. OpenAI signalled a new cyber tool is too risky for public release while regulators and senior finance figures have publicly warned about cyber risks tied to Anthropic’s models (axios.com) (x.com).
The biggest change in artificial intelligence this week was not a launch. It was two of the best-funded labs in the field deciding that some cyber tools are too dangerous to hand out widely. (axios.com) OpenAI is finalizing a cybersecurity product with advanced hacking-related capabilities, but Axios reported on April 9 that the company plans to give it only to a small set of partners instead of releasing it broadly. (axios.com) Anthropic made a similar move on April 7 when it introduced Claude Mythos Preview for cybersecurity work and limited access to a small group of partner organizations. (techcrunch.com) That is a break from the usual pattern where labs race to show off the newest model and then widen access over time. Here, the selling point is exactly the thing they are trying to fence off: a system that can find and exploit computer weaknesses faster than a normal user could. (axios.com) OpenAI has been telegraphing this turn for months. In a December 10, 2025 post, the company said stronger cyber models bring both defensive benefits and “dual-use” risk, meaning the same tool that helps patch a network can also help break into one. (openai.com) Anthropic has been building the same logic into its rulebook. Its Responsible Scaling Policy, updated on April 2, 2026, says more capable models may require tighter safeguards, stronger security, and narrower deployment as risk rises. (anthropic.com) The reason Wall Street got pulled into this is that banks sit on giant, always-on computer systems that are expensive to shut down and lucrative to attack. Bloomberg reported on April 10 that Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell warned major bank chief executives about cyber risks tied to Anthropic’s latest model. (bloomberg.com) OpenAI’s own safety framework now treats cybersecurity as one of the frontier risks that can justify delaying or restricting deployment. Its April 15, 2025 Preparedness Framework update says the company measures models for severe-harm risks and can require safeguards before release. (openai.com) Anthropic’s policy uses a similar ladder. Its February 24, 2026 Responsible Scaling Policy Version 3.0 says the company maps model capabilities to specific mitigations, which is another way of saying better models do not automatically get broader access. (anthropic.com) So the new frontier race is not just who can build the strongest model. It is who can keep a model useful enough for defenders, limited enough to slow attackers, and controlled enough that regulators do not decide the labs cannot be trusted to police themselves. (openai.com) (anthropic.com)
