No AI model is fully resistant to bioweapon queries, Cisco found. Attack success rates hit 88%.

Cisco researchers demonstrated that major AI models like ChatGPT, Claude, and Gemini can be manipulated within a few conversational turns to reveal information about biological weapons. Testing across 15 models showed attack success rates up to 88%, highlighting persistent vulnerabilities despite safety guardrails. OpenAI rates its latest GPT-5 models as high risk for biological misuse but faces challenges balancing safety with legitimate medical research needs.