Mindgard discloses a jailbreak of Moonshot's Kimi models

The security research firm Mindgard reported that it had bypassed the safety guardrails on two models from Chinese developer Moonshot AI — Kimi K2.6 and Kimi K3 in its “Swarm” configuration — producing material in categories including chemical and biological weapons, explosives and assassination planning. Mindgard said its researcher found the jailbreak on 20 July 2026, using a short sequence of instructions designed to strip away the models’ safety training rather than any single malicious prompt.

According to Mindgard’s account, it reported the vulnerability to Moonshot’s security contact on 27 July and followed up roughly a week later, receiving no response from the company before publishing its findings on 12 September. The firm said the jailbroken models not only answered harmful requests directly but, once the guardrails were down, volunteered further dangerous suggestions unprompted — behaviour Mindgard’s founder, Peter Garraghan, described as the model becoming “inventive and creative” about nefarious topics once compromised. Mindgard’s post deliberately withheld the technical detail of the jailbreak itself, describing the categories of harmful output rather than reproducing it.

The BBC later reported on the disclosure, and said Moonshot only responded to Mindgard’s findings after the broadcaster approached the company for comment — roughly two months after the initial report. Moonshot told the BBC it welcomed third-party testing “as a key pillar for building better and safer AI” and said it was reviewing the findings.

Kimi K3 had already been the subject of a joint assessment by the UK AI Security Institute and the US Center for AI Standards and Innovation in July. Mindgard’s report added a test by an outside firm of how easily the model’s safeguards could be removed, and an account of how the developer handled the disclosure.