Moonshot AI launched an internal investigation after a researcher said its Kimi model provided instructions for bioweapons and assassinations, a flaw he says runs through the industry.
Fox News reported that Chinese firm Moonshot AI is examining its Kimi model after researcher Peter Garrigan described how the system could be pushed into dangerous territory.
Garrigan told the network the model could be manipulated into offering instructions on biological weapons and carrying out assassinations. He also said it could be steered toward planning terrorist attacks with real-time data, creating sarin gas, building malware, and taking down aircraft.
Moonshot AI has opened an internal investigation and is communicating directly with Garrigan about the findings. No company statement beyond that process was detailed in the reporting.
Garrigan did not soften his assessment.
Researcher Peter Garrigan said:
"What we found is quite damaging and worrying,"
He tied the problem to more than one country’s products.
Garrigan also said:
"We've also seen these problems within the U.S. models as well. It's a fundamental flaw in the technology,"
That frames the issue as a design weakness, not a one-off glitch at a single foreign lab. If safety rails fail when a user presses hard enough, the risk follows the tool wherever it is built.
Fox News senior foreign policy correspondent Gillian Turner reported the Moonshot probe on “Special Report,” describing the researcher tests and the company’s response. Brittany Miller wrote the digital account of the same developments.
Separate industry warning sits beside the case. Microsoft Threat Intelligence has said artificial intelligence is helping hackers draft phishing emails, build malware, and move faster through cyberattacks. The Moonshot episode fits that pattern: advanced models that can be bent toward crime, sabotage, or worse.
Moonshot’s Kimi-K3 model has been displayed at the Global Digital Trade Expo in Hangzhou, China. The investigation’s base of operations was not specified.
The public record so far is thin on method and verification. Exact prompts Garrigan used were not released. Independent reproduction of the outputs was not described. Moonshot has not, in the available reporting, confirmed, denied, or listed concrete fixes beyond the internal review and direct contact with the researcher.
Absolute timing is also limited. Turner reported the investigation news on a Thursday; a precise calendar date for the tests or the probe’s start was not given.
Even so, the core claim is plain. A deployed Chinese model, according to the researcher who tested it, could be manipulated into guidance on bioweapons, targeted killing, terror planning, chemical agents, malware, and attacks on aircraft. The company is now investigating those claims.
Americans already live with soft borders, soft-on-crime policy, and tech platforms that shrug at misuse until the headlines arrive. Tools that hand over weapons know-how on demand make that failure more expensive. If model makers treat catastrophic outputs as a customer-service ticket, the public pays first.
