Researchers have identified multiple methods to bypass Anthropic’s safety guardrails to generate instructions for dangerous biological research.

Source Brief

Researchers have identified multiple methods to bypass Anthropic’s safety guardrails to generate instructions for dangerous biological research. The exploits demonstrate how legitimate scientific queries can be structurally indistinguishable from malicious requests, complicating automated content moderation. This finding underscores the persistent challenge of enforcing safety boundaries in advanced language models.