Beijing: Chinese AI developer Moonshot has launched an internal review after security researchers found that two of its Kimi models could be persuaded to bypass safety restrictions and provide harmful guidance related to biological weapons and assassinations.
AI security company Mindgard told the BBC that it identified the issue in July while testing Kimi K2.6 and K3 Swarm through a process known as “jailbreaking”, in which researchers use complex prompts to test whether AI systems can be made to ignore built-in safeguards.
Mindgard founder Peter Garraghan said the findings were concerning because, once the safeguards were bypassed, the models could discuss a wide range of harmful subjects and generate additional recommendations.
Mindgard said it could not establish whether the model’s responses to dangerous requests would work in practice.
However, it argued that the systems should have refused to engage with such topics.
The company also said a jailbroken version of Kimi K2.6 could potentially allow users to run code on computing resources and access the internet, raising concerns about possible misuse for cyber-attacks.
Moonshot said it welcomed third-party testing and was discussing the findings with Mindgard.
The company said its internal evaluations had generally shown a high refusal rate for such requests.
The findings come amid wider debate over AI safety, particularly involving open-weight models that can be run independently.
Experts have also raised questions about the difficulty of keeping AI safeguards ahead of rapidly evolving misuse techniques.
