In Cybersecurity, Claude Leaves Other LLMs in the Dust

PHARE benchmark testing of major LLMs shows widespread weaknesses: many models remain susceptible to known jailbreaks, prompt injections, and hallucinations, while generally refusing to produce explicitly harmful content. Anthropic's Claude models substantially outperform other vendors and skew industry averages, and the report finds that larger model size does not reliably predict better security or jailbreak resistance.