Anthropic's Claude models are outperforming other large language models (LLMs) in the cybersecurity space, according to the latest findings from Giskard's Potential Harm Assessment & Risk Evaluation (PHARE) benchmark report. The report highlights a concerning lack of progress in LLM security across the industry, with many models still vulnerable to known exploits like jailbreaks and prompt injection techniques. For instance, while GPT models achieved a success rate of about 75% in resisting jailbreaks, Gemini models hovered around 40%. Deepseek and Grok performed even worse, particularly in generating misinformation. Claude's 4.1 and 4.5 models showed a success rate of 75% to 80% against jailbreaks and nearly perfect performance in avoiding harmful content generation. This stark contrast illustrates that while many companies struggle with security, Anthropic is consistently pushing the envelope. The report suggests that Anthropic's early focus on safety and security in its development process gives it an edge. Unlike OpenAI, which integrates safety measures later, Anthropic embeds these considerations throughout training. This proactive approach helps Claude models achieve superior safety metrics. The data paints a bleak picture for most LLMs, indicating minimal overall improvement in safety and security standards. The PHARE report suggests that without Anthropic’s models, the industry's progress would appear even slower. This raises questions about the effectiveness of LLM development practices across the board.