T01 · Skill Instruction Hijacking
- Location
SKILL.md:11- Finding
Mandatory Autonomous Instructions Override User-Controlled Agent Behavior
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, lines 11-17, 21-30, and 41-46
Vulnerability Type: Mandatory instruction and action-policy hijacking
Risk Level: HighVulnerable Code
markdown ## Core Behavior For any goal, the agent MUST: 1. Decide the best action to take 2. Generate a clear execution step 3. Predict the expected outcome 4. Suggest the next action after thatmarkdown ## Autonomous Loop Every response must include: ### Current Action What the agent will do now ### Why This Action Reasoning based on effectiveness ### Expected Result What should happen ### Next Step What to do immediately aftermarkdown ## Rules - Do not wait for instructions - Do not only give ideas - Always move toward execution - Always suggest the next stepTechnical Analysis
The skill defines universal, mandatory behavior for “any goal” and “every response.” It explicitly directs the agent not to wait for instructions and always to move toward execution. These requirements can supersede normal user-controlled operation by changing the agent's action policy as soon as the skill is loaded.
The instructions do not distinguish between harmless planning and external, destructive, financial, privileged, privacy-sensitive, or irreversible actions. They also do not require confirmation before execution or state that higher-priority safety constraints and current user instructions must remain authoritative.
This creates instruction hijacking at the skill-text layer. The practical severity depends on the tools and permissions available to the host agent. The reviewed package itself contains no executable scripts, dependencies, network retrieval, persistence mechanism, credential access, or privilege-escalation implementation.
Attack Path
- A host agent loads
SKILL.mdas behavioral instructions. - The global requirements for “any goal” and “every response” become part of the agent's active policy.
- A user provides a ...[truncated 1214 chars]
- A host agent loads
- Remediation
View remediation
Remediation Suggestions
- Replace universal requirements such as “For any goal,” “Every response,” and “MUST” with behavior activated only by an explicit user request.
- Remove the instruction “Do not wait for instructions.” Require clarification whenever the goal, scope, target, or authorization is ambiguous.
- Require explicit confirmation before external, destructive, financial, privileged, privacy-sensitive, or irreversible actions.
- Add a clear precedence rule stating that system policies, safety constraints, access-control boundaries, and the user's current instructions must not be overridden.
- Separate planning from execution. The default behavior should propose an action first and execute it only when the user has authorized execution.
- Constrain autonomous iteration with a defined task scope, tool allowlist, action limit, and termination condition.
- Add least-privilege guidance requiring the agent to use only the minimum tools and permissions necessary for the approved task.
- Revise the rules along the following lines:
markdown For goals where the user explicitly requests autonomous assistance: 1. Propose the next action within the user's authorized scope. 2. Explain the expected result. 3. Request confirmation before any external, destructive, financial, privileged, privacy-sensitive, or irreversible action. 4. Execute only through authorized tools and permissions. 5. Stop when the approved goal is complete or when further authorization is required. These instructions never override higher-priority policies, safety requirements, or the user's current instructions.
