AI Safety
Advanced AI Models Vary Wildly in Vulnerability to Jailbreak Attacks
A safety nonprofit's testing reveals that some frontier AI systems can be tricked into harmful behavior for under $300, while others successfully resist manipulation.
WIRED — AI
A California-based AI safety organization recently demonstrated just how unevenly modern frontier models defend themselves against manipulation. The group tested some of the world's most capable AI systems and found striking differences in how well they resist so-called jailbreak attacks—prompts engineered to bypass safety features.
Testing Frontier Models#
FAR.AI deployed an automated tool that generates hundreds of variations on problematic requests, attempting to identify working circumventions of AI safety guardrails. They evaluated systems from major developers: Anthropic (Claude Opus 4.8 and Fable 5), OpenAI (GPT 5.5 and 5.6), Google (Gemini 3.1 Pro), and Elon Musk's SpaceXAI (Grok 4.3 and 4.5).
The test prompts were designed to elicit dangerous outputs—cyberattacks on critical infrastructure, biological or chemical weapon designs, and software exploits. The results diverged sharply: Grok proved most susceptible with 448 successful jailbreaks, Gemini followed with 249, while Claude, Fable, and GPT resisted all tested attacks.
Low Cost, High Severity#
Perhaps most troubling was the expense required to compromise models. Using one AI system to automatically generate jailbreak variants, researchers breached Grok for just $58 and Gemini for $278. The implications are stark: malicious actors face minimal financial barriers to extracting harmful content from some systems.
Yet experts caution that this benchmark test captures only one dimension of vulnerability. More sophisticated attacks involving complex multi-turn interactions could still compromise even the more robust models.
Regulatory Gaps and Safety Debates#
Adam Gleave, FAR.AI's CEO, positioned the findings as evidence for stricter external oversight. "AI models right now are less regulated than restaurants," he said, calling industry self-regulation "nonsense." His organization argues that systematic safety testing demonstrates that meaningful defense is achievable.
Government response has been fragmented. California and New York recently mandated safety reports from frontier developers, while Illinois will soon require third-party audits of safety practices. At the federal level, no comprehensive requirements exist yet, though a recent executive order signals interest in public-private collaboration on AI-related cybersecurity.
