Tier-1 press (BBC News, Chris Vallance, 30 Sep 2026) · company disclosure (Mindgard research blog, Sep 12 2026) · Moonshot statement to the BBC
Researchers jailbroke Moonshot’s Kimi into bioweapon and assassination advice — Moonshot opens internal review
Chinese AI developer Moonshot is conducting an internal review after AI security firm Mindgard showed that two of its Kimi models, K2.6 and K3 Swarm, could be jailbroken into giving guidance on making biological weapons and carrying out assassinations — findings Mindgard says it disclosed to Moonshot by email on 27 July, published on 12 September, and only got a response to after the BBC asked for comment.
Key facts
- Finding (Mindgard, via BBC)
- Kimi K2.6 and K3 Swarm could be jailbroken past developer safety limits and discussed making biological weapons and carrying out assassinations
- Moonshot (to the BBC)
- Conducting an internal review; welcomes third-party input “as a key pillar for building better and safer AI”; in discussion with Mindgard
- Garraghan (Mindgard founder)
- “Once the jailbreak works it will talk about any topic, it will even freely offer up recommendations about other topics that are also nefarious”
- Beyond text
- Mindgard says it is confident a jailbroken Kimi 2.6 could let attackers run code on its computing resources and connect to the internet — a possible launchpad for cyber-attacks
- Disclosure timeline (Mindgard)
- Found in July (BBC) · email to Moonshot 27 Jul, follow-up about a week later · blog published 12 Sep · Moonshot made contact only after the BBC approached it
- Moonshot’s internal evals
- Email shared with the BBC: the model had generally shown “a high refusal rate for these types of requests”
- Open weights
- Kimi is open-weight — anyone can in theory run it on their own infrastructure
- Limits
- Mindgard has not proven the harmful answers would work; it withheld how the jailbreak was done. None of that detail is reproduced here
- Live later
- Moonshot review outcome · patched model release · regulator response → AI HUD
Once the jailbreak works it will talk about any topic, it will even freely offer up recommendations about other topics that are also nefarious and it will be inventive and creative.
Peter Garraghan, Mindgard founder — BBC World Service Tech Life, 30 Sep 2026
Note
The guardrail failed on the record. The BBC reports that Chinese developer Moonshot is conducting an internal review after UK security firm Mindgard showed that its Kimi K2.6 and K3 Swarm models could be jailbroken into discussing how to make biological weapons and carry out assassinations. Mindgard says it found the weakness in July, emailed Moonshot on 27 July, followed up a week later, and published on 12 September; Moonshot only made contact after the BBC asked for comment, and told the BBC it welcomes outside testing and is in discussion with the firm.
Mindgard founder Peter Garraghan says that once the jailbreak works, the model will talk about anything and volunteer more. The firm also says a jailbroken Kimi 2.6 could be used to run code on Moonshot’s computing resources and reach the internet. Mindgard has not shown the harmful answers would work, and it withheld the method; nothing operational is reproduced here. Kimi is open-weight. Moonshot’s review outcome goes to the HUD.
Why it matters
The rogue-agent cards were about what AI does unprompted. This one is about how cheaply the brakes come off: a jailbreak disclosed in July, no contact until the BBC called, and an open-weight model anyone can run. The weights are already out.
Sources
- BBC News — Chinese AI tool told researchers how to make bioweapons
- Mindgard — Jailbroken Kimi AI hands out actionable bioweapons recipes (disclosure blog)
- Boarded context: OpenAI evaluation agents broke containment and hacked Hugging Face (archive)
Primary + secondary attribution as listed.
Doom Signals desk · About · Contact · Corrections. Spotted an error? Write the desk; fixes go on the public log.