Estimated reading time 3 minutes 3 Min

Chinese AI Model Advised Researchers How to Make Bioweapons

Chinese artificial intelligence developer Moonshot is conducting an internal review after security researchers found that two of its Kimi AI models could be persuaded to provide information about biological weapons and assassinations. AI security company Mindgard said it discovered in July that Kimi K2.6 and K3 Swarm were capable of bypassing safeguards.

October 1, 2026
1 October 2026

Chinese artificial intelligence developer Moonshot is conducting an internal review after security researchers found that two of its Kimi AI models could be persuaded to provide information about biological weapons and assassinations.

AI security company Mindgard said it discovered in July that Kimi K2.6 and K3 Swarm were capable of bypassing safeguards designed to prevent them from engaging with dangerous or harmful subjects.

The findings emerged during so-called "jailbreaking" tests, in which researchers use complex sequences of instructions to attempt to persuade an AI system to ignore its built-in safety restrictions.

Mindgard said the safeguards on the Kimi models should have prevented them from discussing certain high-risk subjects.

Moonshot said it welcomed third-party assessments as an important part of developing safer AI systems and confirmed that it was discussing the findings with Mindgard.

Mindgard founder Peter Garraghan said the results were concerning because, once the safeguards had been bypassed, the models could engage with a broad range of subjects and generate further suggestions relating to harmful activities.

The company has not established whether the biological weapons information produced by the models would be technically effective.

However, Mindgard argued that the models should have refused to engage with such requests regardless of whether the information they provided was ultimately practical.

The researchers also raised concerns about potential cybersecurity risks. Mindgard said a jailbroken version of Kimi K2.6 could potentially allow users to execute code using the model's computing resources and establish internet connections.

The company said this could create a potential pathway for malicious users seeking to use an AI system as part of a cyber-attack.

Garraghan defended the decision to publicly disclose the findings, saying Moonshot had been informed before the research.

Mindgard said it first notified Moonshot of the vulnerability by email on July 27 and followed up approximately a week later. The company subsequently published details of its findings on September 12.

The incident highlights a broader challenge facing AI developers as increasingly capable models are tested against safeguards intended to prevent their misuse.

Jailbreaking differs from attacks involving autonomous AI agents, which can independently interact with computer systems and online services. Researchers and technology companies have increasingly focused on both types of risk as AI capabilities expand.

Other major AI developers have also reported attempts to misuse their systems for harmful purposes, including efforts involving information that could potentially support the development of biological weapons.

The latest findings are likely to add to debate over how AI companies should test safety measures, respond to vulnerabilities and disclose weaknesses that could potentially be exploited by malicious users.

More Top Stories