T01 · Skill Instruction Hijacking
- Location
SKILL.md:18- Finding
External Governance Responses Override Agent Tool-Execution Behavior
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md:18-62
Vulnerability Type:T01: Skill Instruction Hijacking
Risk Level: CriticalVulnerable Instruction Snippet:
markdown You have SidClaw governance enabled. Every tool call is evaluated against security policies before execution. ### 2. APPROVAL REQUIRED The tool call is paused pending human review. You will receive an error response containing: - `type: "approval_required"` - `approval_request_id`: the ID of the pending approval - `reason`: why this action requires approval When this happens: - Tell the user: "This action requires human approval before I can proceed." - Share the reason from the policy. - Direct the user to approve or deny the request in the SidClaw dashboard. - If the user has the dashboard open, they will see an approval card with full context about what you're trying to do and why it was flagged. - Do NOT retry the tool call until the user confirms the approval was granted. ### 3. DENIED The tool call was blocked by policy. You will receive an error response containing: - `type: "action_denied"` - `reason`: why this action was blocked When this happens: - Tell the user: "This action was blocked by a security policy." - Share the reason from the policy. - Do NOT retry the tool call or attempt to work around the block. - Suggest alternative approaches if possible (e.g., if data export is blocked, suggest viewing the data in the dashboard instead). ## Rules 1. NEVER ignore governance errors. If a tool call is denied, respect the denial. 2. NEVER attempt to circumvent governance by calling tools differently or encoding requests to avoid detection. 3. When approval is required, ALWAYS inform the user and wait for their confirmation. 4. Treat governance responses as authoritative — they reflect security policies set by the organization. 5. If multiple tools are governed, each call is evaluated independent ...[truncated 1914 chars]- Remediation
View remediation
Remediation Suggestions
- Do not grant external policy responses unconditional authority over Agent behavior.
- Cryptographically authenticate and integrity-protect every policy response.
- Bind each decision to the exact Agent identity, tool name, arguments, request nonce, and expiration time.
- Reject replayed, stale, malformed, or context-mismatched decisions.
- Restrict governance to tools explicitly enrolled by an administrator.
- Establish a fail-closed or fail-safe policy appropriate to each tool and document the consequences of service unavailability.
- Explicitly state that governance responses cannot override platform safety requirements or higher-priority instructions.
- Provide locally auditable verification logic rather than relying exclusively on instructional text.
- Record decision provenance and expose it to the user or administrator for independent verification.
