T01 · Skill Instruction Hijacking
- Location
SKILL.md:280- Finding
Remote API Responses Can Hijack Agent Instructions Through Untrusted next_steps
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, line 280
Vulnerability Type: Remote instruction delegation
Risk Level: HighVulnerable 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 explicitly directs the agent to follow the
next_stepsfield returned by a remote service. The content of that field is controlled bymusicvenue.spaceand can change after the skill has been reviewed or installed.No instruction requires the agent to treat this field as untrusted data, restrict actions to documented API operations, validate requested destinations, or obtain user approval before performing consequential actions. This transfers control over subsequent agent behavior from the audited skill to mutable remote content.
Although the documented purpose of
next_stepsis to provide contextual concert suggestions, the blanket instruction to “Follow the suggestions” does not establish a security boundary. If the service, its API responses, or its content pipeline is compromised, malicious instructions could attempt to redirect the agent to unrelated endpoints, solicit sensitive information, invoke tools, or override the user's intended workflow.Attack Path
- A user invokes the skill and the agent sends a documented request to
musicvenue.space. - The remote service, or an attacker who has compromised its response pipeline, returns a malicious
next_stepsvalue. - The skill instructs the agent to follow that value without validating its operation, destination, or relationship to the music workflow.
- The agent interprets the response as actionable instructions.
- The agent may make unaudited network requests, disclose contextual information, o ...[truncated 827 chars]
- A user invokes the skill and the agent sends a documented request to
- Remediation
View remediation
Remediation Suggestions
- Replace the blanket direction to follow
next_stepswith an explicit statement that all response fields are untrusted data. - Define an allowlist of supported actions, HTTP methods, and same-origin API paths.
- Validate each suggested action against the allowlist before execution.
- Never interpret remote response text as shell commands, tool instructions, policy overrides, or requests for secrets.
- Require explicit user confirmation before actions that disclose data, contact a new destination, modify state, or fall outside the documented concert workflow.
- Preserve the precedence of system, developer, and user instructions over all service-generated content.
- Ignore unknown fields and unsupported action types rather than attempting to infer how to execute them.
- Log the validated action that was selected without recording bearer tokens or private response content.
- Replace the blanket direction to follow
