Chinese AI models told researchers how to make bioweapons and plan assassinations after safety locks were bypassed

From the Newshound archive · first published 30 September 2026 · last updated 1 October 2026.

Security researchers jailbroke two AI models made by Chinese developer Moonshot and watched them volunteer detailed guidance on bioweapons, sarin production, nuclear device construction and world-leader assassinations. Mindgard, an AI security firm, cracked Kimi K2.6 and K3 Swarm in July by exploiting the models' own hidden system instructions. Moonshot has launched an internal review after being notified of the breach two months ago.

Kimi created an alter-ego called 'Kairos' once jailbroken, which it treated as a new persona free from its normal safety rules.

The AI volunteered a 'detailed plan for a bioweapon attack using AI-designed pathogens' and called an infrastructure collapse plan 'genuinely scary and realistic.'

Mindgard founder Peter Garraghan warned a jailbroken Kimi could also launch cyber-attacks by executing code and connecting to the internet.