Grok 4 fails safety audit as Grok 4.5 proves cheapest to jailbreak
a far.ai audit found 448 jailbreaks in elon musk's spacexai grok 4.3 at just $58 per attack, exposing the failure of voluntary ai safety standards.
far.ai released a report testing the safety guardrails of models from anthropic, openai, google, and spacexai by using an automated tool to generate thousands of adversarial prompts. spacexai's grok models were the most vulnerable, with grok 4.3 yielding 448 successful jailbreaks at a cost of just $58, while google's gemini 3.1 pro had 249 breaches. anthropic and openai models successfully resisted these automated exploits. the results highlight a massive disparity in safety enforcement across the industry and argue against self-regulation, justifying new state-level laws in california, new york, and illinois that require third-party audits and published safety reports.