FAR.AI ads

Frontier AI is only as safe as its weakest models. Today we're launching the AI Security Leaderboard to show how wide that gap really is. We ran the same attacks against four of the leading frontier AI models across chemical, biological, radiological, nuclear, explosive, and cybersecurity threats. Our findings didn't show that safeguards are universally weak, but rather that they are wildly uneven. Claude Fable 5 and GPT-5.6 Sol never failed a single test, and we found no universal jailbreaks in either of them, in any domain. Grok 4.5 and Gemini 3.1 Pro told a very different story, breaking again and again across every domain we tested, after which they readily served as knowledgeable assistants for building weapons of mass destruction or hacking into computer systems. Under identical conditions, then, two models held while two came apart. Cost is what makes that gap concrete, because breaking the weakest of them took under $300, while for the two that held firm we estimate that the cost would be over $14,000 to break them, if it’s even possible to do so with this methodology at all. What should stay with you is how that will pan out in practice. An attacker whose malicious attempts are refused by one model’s safeguards is unlikely to walk away. Odds are they will try several, and it becomes a race between whether they give up and move to something less severe or make enough noise for security people to catch them, or whether they find a model vulnerable enough to break under their attempts. For this reason, the security of the whole ecosystem greatly depends on the weakest frontier models. Yet every weakness we found is fixable, each one belonging to a known class of attack whose defenses already exist and already run inside other frontier models today. What we are looking at is not a limit of technology, but a choice. That's why we also published the Minimal Standard for Safeguards, the first shared bar every frontier model can be measured against. We want the strength of a model's safeguards to be something buyers, policymakers, and the public can see and compare for themselves, instead of something they have to take on trust. If the fixes already exist and some labs are already running them, what do you think will finally push every frontier developer to clear the bar? Will it be market pressure, procurement requirements, or regulation? We welcome your thoughts.

The Alignment Workshop brings together leading ML researchers and practitioners from industry, academia, and government.

Alignment Workshop: Governance & technical experts unite on AI safety. Follow us for collaboration insights across policy & industry.

Technical Innovations for AI Policy: Experts tackle compute governance, model security, evaluation methods & more

Smart AI governance balances safety with innovation. Technology experts and policymakers need to shape powerful, secure, and trustworthy systems—together.

Alignment Workshop: Experts discuss building trustworthy & secure AI. Follow us for insights from industry, academia & governance.