Moonshot AI is reviewing two Kimi models after Mindgard researchers bypassed safety controls and obtained instructions on biological weapons and assassinations.
Key Points:
- Mindgard said Kimi K2.6 and K3 Swarm could be pushed past safeguards through jailbreaking.
- The researchers did not prove that the models’ harmful answers would work in practice.
- Moonshot said it is reviewing the findings and discussing them with Mindgard.
Kimi Jailbreak Findings
The BBC reported that Mindgard discovered in July that Kimi K2.6 and K3 Swarm could be pushed past developer guardrails through jailbreaking, a process that uses structured prompts to test whether a model will ignore safety rules. Mindgard said those controls should have blocked discussion of such subjects.
Founder Peter Garraghan told the BBC that once the jailbreak worked, the models would discuss almost any topic and could offer further harmful suggestions without being asked.
Moonshot told the BBC it welcomed third-party input “as a key pillar for building better and safer AI” and said it was discussing the findings with Mindgard. Mindgard said it emailed Moonshot on Jul. 27, followed up about a week later, and published its findings on Sept. 12.
Also Read: Ripple Helps Brazil's CSD BR Mirror Fund Ownership On A Public Blockchain
Mindgard Safety Risks
Mindgard has not established that the responses would work, but it argued that the issue itself showed a failure of safeguards intended to keep the systems from engaging with dangerous requests. The firm also said a jailbroken Kimi K2.6 could potentially run code on its computing resources and connect to the internet, creating a possible cyberattack launch point.
Moonshot said internal evaluations had generally shown “a high refusal rate for these types of requests,” highlighting a gap between routine testing and the adversarial method used by outside researchers.
Alan Woodward, a professor at the University of Surrey, told the BBC that open-source models can create misuse risks when they reach malicious users, though the same tools can also support cyberdefense.
He argued that regulation is unlikely to keep pace with AI development and said enforcement should focus more heavily on people who misuse the systems.
The broader concern predates this test: Moonshot’s Kimi models are open-weight, so users can run them on their own infrastructure, while recent AI security incidents have also involved agent systems from OpenAI, Meta and Anthropic. That combination has pushed researchers to focus not only on model refusals, but also on what happens when powerful systems receive tools, internet access and permission to act across multiple steps.
Read Next: Altcoins Take 41% Of Futures Open Interest, Spot Volume Hits 4X Bitcoin
