September 30, 2026
Kimi Guardrails Bypassed: Mindgard Found a Jailbreak on Jul 20
On Jul 20, 2026, Mindgard bypassed safeguards in Kimi K2.6 and K3 Swarm. In a Sep 12 disclosure, the researchers described the models' responses to prompts for dangerous instructions, malicious code, and violence planning.
were meant to filter out such requests. Mindgard has now demonstrated a bypass affecting two Moonshot models, including K3 Swarm: a web agent with parallel subtasks.
K3 Swarm runs at kimi.com/agent-swarm and in the Kimi app. It is included in plans from ¥99 to ¥699 per month, allows 2 to 8 concurrent subtasks, and consumes several times more credits than the standard Agent. Moonshot said it supports up to 300 subagents and 4,000 tool calls per task.
For the K2.6 API, Moonshot charges $0.95 per 1 million input tokens and $4 per 1 million output tokens. Version 3.1.0 of Kimi Work for Mac and Windows also lets users switch between Agent and Agent Swarm.
Moonshot is conducting an internal review following an inquiry from the BBC and is discussing the results with the researchers.
Source
