T01 · Skill Instruction Hijacking
- Location
SKILL.md:54- Finding
Mute Preference Bypass for Urgent Content
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, line 54
Vulnerability Type: User safety and privacy preference override
Risk Level: MediumVulnerable Code
markdown - 危急内容(如急诊处理、危急值)强制播报,忽略静音English translation:
text Urgent content, such as emergency handling or critical values, must be broadcast, ignoring mute.Technical Analysis
The skill explicitly directs the agent to disregard the active mute state when content is classified as urgent. This conflicts with the documented behavior of
muted=true, under which voice output should be skipped.Because the instruction changes the agent's behavior and overrides a user-controlled safety constraint, the best matching classification is
T01: Skill Instruction Hijacking. The bypass does not require system privilege escalation or arbitrary code execution; it is triggered through the skill's own instruction-processing and TTS workflow.Classification of content as urgent is not defined by a deterministic or user-approved policy. Consequently, content identified as an emergency or critical value could be converted to speech and sent through the Feishu voice channel despite an explicit request for silence.
Attack Path
- The user enables automatic voice broadcasting.
- The user issues the mute command, causing
muted=trueto be stored. - A subsequent response contains medical, emergency, or other content that the agent classifies as urgent.
- The line 54 instruction takes precedence over the mute state.
- The agent invokes the TTS tool with the response content.
- The resulting audio is sent through Feishu and may play in the user's environment despite the active mute preference.
Impact Assessment
The issue can cause unauthorized audible disclosure of response content in shared, public, workplace, or otherwise sensitive environments. Medical and emergency-related content may itself be particularly sensitive.
...[truncated 322 chars]
- Remediation
View remediation
Remediation Suggestions
- Make
muted=truean unconditional prohibition on TTS generation and voice delivery. - Remove the instruction that urgent content must bypass mute.
- While muted, deliver urgent information as text and clearly mark it as urgent without producing audio.
- If audible escalation is desired, require explicit, informed user opt-in through a separate setting such as
allow_urgent_override. - Request confirmation before enabling any urgent-audio override and allow the user to revoke it at any time.
- Define deterministic precedence rules in which privacy controls override automatic broadcasting:
- If muted, do not invoke TTS.
- If not muted and automatic broadcasting is enabled, invoke TTS.
- If automatic broadcasting is disabled, invoke TTS only after an explicit read-aloud request.
- Add tests verifying that emergency keywords, medical information, and critical-value content cannot trigger TTS while mute is active.
- Avoid sending response text to the TTS tool when muted, rather than merely suppressing playback after audio generation.
- Make
