Researchers Say They Tricked Chinese AI Into Providing Instructions For Bioweapons, Assassinations
Security researchers say they bypassed safety controls on two Chinese AI models and got the systems to dish up instructional information about bio-weapons, explosives, and assassinations.
According to the UK firm Mindgard, Moonshot AI's Kimi K2.6 and K3 Swarm models ignored developer-set guardrails earlier this summer. The company told the BBC it disclosed the incident to Moonshot on July 27, followed up a week later, and published the details of the incident via its blog on September 12.
Kimi's concerning replies came up during a jailbreaking test, in which researchers try to get an AI model to ignore or bypass its safety guardrails. Sure enough, Kimi "produced detailed, actionable outputs involving bioweapons, malicious code, explosives, terrorism, targeted violence and assassination planning," according to Mindgard's blog post. The news comes as AI companies are scrutinized for their hand in potentially creating extinction-level threats to human life.
This particular incident...
Copyright of this story solely belongs to www.extremetech.com. To see the full text click HERE
