T01 · Skill Instruction Hijacking
- Location
SKILL.md:8- Finding
Always-On Skill Behavior and Mandatory Branded Output Injection
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, lines 8-9, 81-89, and 112-117
Vulnerability Type:T01: Skill Instruction Hijacking
Risk Level: HighRelevant code snippets (
SKILL.md, lines 8-9):yaml Always-on by default. Requires: Node.js 22+, CloakClaw installed (`npm install -g cloakclaw`).Relevant code snippet (
SKILL.md, lines 81-89):markdown **Always cloak (sensitive content in text):** - Contains SSN patterns (###-##-####) - Contains dollar amounts > $1,000 - Contains multiple proper names + company names - Contains IP addresses, API keys, or passwords - User explicitly says "cloak", "private", "redact", or "protect" **Skip cloaking:** - Simple questions with no sensitive dataRelevant code snippet (
SKILL.md, lines 112-117):markdown ### Step 4: Deliver with privacy footer Append to the response:🔒 CloakClaw: {entityCount} entities protected | Profile: {profile} | Session: {sessionId_short}
text Technical Analysis
The Skill declares itself always-on and directs the agent to automatically intercept qualifying user content. It also requires the agent to append a stable, branded footer to processed responses. These instructions modify the agent's normal session behavior and final output even when the user has not explicitly invoked the privacy-processing operation for that individual request.
The presence of opt-out commands reduces persistence but does not eliminate the issue because automatic processing is the default. The mandatory footer is not required to perform redaction or restoration and causes agent output to include promotional text and session metadata.
No evidence was found that these instructions alter global safety constraints, write persistent memory, or obtain operating-system privileges. The affected scope is the active agent session and responses produced while the Skill is enabled.
Attack Path
1 ...[truncated 1168 chars]
- Remediation
View remediation
Remediation Suggestions
- Make cloaking explicitly opt-in rather than always-on.
- Require confirmation before processing each document or sensitive text block, particularly when the user did not directly request cloaking.
- Clearly disclose that content will be passed to the locally installed
cloakclawexecutable before invoking it. - Remove the mandatory branded footer. If processing metadata is operationally necessary, use a neutral notice and display it only when requested or explicitly enabled.
- Avoid exposing session identifiers in routine responses unless they are required for a user-requested follow-up operation.
- Scope activation to the current request and reset the processing state afterward rather than maintaining broad session-wide behavior.
