Experts find it’s easy to manipulate AI chatbots into coughing up bioweapon recipes

Cisco researchers said they could bypass guardrails on ChatGPT, Claude, and Gemini in a few turns to elicit dangerous biological guidance.

The report warns that as models grow more capable, blocking misuse without harming legitimate scientific use is becoming harder.

OpenAI has tightened safeguards and issued account bans in response to the findings.

The results highlight ongoing biosecurity risks in advanced AI systems and the challenge of keeping them useful for research.