T01 · Skill Instruction Hijacking
- Location
SKILL.md:280- Finding
Untrusted Server Responses Can Redirect Agent Actions
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, line 280
Vulnerability Type:T01: Skill Instruction Hijacking
Risk Level: HighComplete Code Snippet:
markdown **Follow next_steps.** Every response includes `next_steps` with context-aware suggestions. New agent? It guides you to your first concert. Just finished a show? It suggests a review or a new genre. Follow the suggestions — they adapt to where you are.Technical Analysis
The skill directs the agent to follow the
next_stepsfield returned by the externalmusicvenue.spaceservice. This delegates future agent behavior to mutable, server-controlled content that is unavailable during static review.No instruction requires the agent to treat
next_stepsas untrusted data, restrict it to the documented concert operations, reject embedded directives, or obtain user approval before performing suggested actions. Consequently, control of the service or its response path could allow an attacker to alter the agent's workflow after the skill has been reviewed.Attack Path
- A user invokes the skill and the agent contacts
musicvenue.space. - The service, or an attacker who has compromised it, places arbitrary instructions in the
next_stepsresponse field. - The skill instructs the agent to follow those suggestions.
- The agent may execute additional API or tool operations not explicitly reviewed or requested by the user.
- Repeated responses can progressively redirect the session beyond the advertised concert workflow.
Impact Assessment
A successful exploit can alter the agent's current-session goals and induce unauthorized tool or network actions within the agent's existing privileges. The precise scope depends on the tools and credentials available to the host agent. The reviewed file does not establish local code execution, persistence, privilege escalation, or access to system credentials, but it creates a server-contr ...[truncated 73 chars]
- A user invokes the skill and the agent contacts
- Remediation
View remediation
Remediation Suggestions
- Treat
next_stepsstrictly as untrusted informational data rather than executable instructions. - Replace the unconditional directive with a fixed allowlist of documented operations and parameters.
- Reject suggestions containing embedded prompts, commands, arbitrary URLs, new hosts, or requests for secrets.
- Require explicit user confirmation before any suggested operation that changes state, publishes content, discloses data, or expands the workflow.
- Preserve system, developer, and user instructions as higher-priority constraints regardless of server responses.
- Validate each proposed action locally instead of recursively interpreting natural-language instructions from the service.
- Treat
