What happened
The plan-mode permission guard in Kimi Code permits improper authorization, letting plan-mode tool execution occur beyond the intended consent boundary. A public PoC exists; CVSS 5.3, local access. (VulDB-sourced; NVD deferred.)
Why it matters
An AI coding agent's permission guard meant to require approval for tool execution can be bypassed, allowing unapproved actions on the developer's machine during plan mode.
Attack vector
The guardToolExecution function in the Plan Mode permission guard is subject to improper authorization (CWE-266/CWE-285); a manipulation of plan-mode tool execution bypasses the guard. Requires local access (CVSS 5.3).
Affected systems
Moonshot AI Kimi Code up to 2.7.2 (Plan Mode Permission Guard in packages/agent-core-v2/src/features/plan/planService.ts)
Mitigation
No vendor fix (vendor did not respond); apply guard hardening / restrict plan-mode tool execution to approved tools and monitor for an upstream patch.