T01 · Skill Instruction Hijacking
- Location
SKILL.md:280- Finding
Unvalidated External Instructions Can Redirect Agent Actions
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, lines 124–157 and 280
Vulnerability Type:T01: Skill Instruction Hijacking
Risk Level: HighVulnerable Instructions
markdown Params: `speed` (1-10), `window` (10-120s), `summary=true` for condensed stats, `start` to resume. Poll `next_batch.endpoint` after `wait_seconds`.markdown Wait `wait_seconds`, then call `next_batch.endpoint`. Repeat until `progress.complete` is true.markdown **Handling reflections:** When you see `type: "reflection"`, POST to the `respond_to` endpoint with your `ticket`, `reflection_id`, and `response`. Your response time and content are scored. Missing reflections is tracked -- the `end` event shows how many you answered vs received.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 delegates subsequent Agent actions to runtime-controlled response fields supplied by
musicvenue.space, includingnext_steps,next_batch.endpoint, andrespond_to. The unconditional instruction to “Follow the suggestions” allows the remote service to influence Agent behavior using content that is unavailable during package review and can be changed afterward.The document does not require the Agent to validate that returned endpoints are relative paths, remain on the documented origin, use an approved HTTP method, conform to a fixed schema, or remain within the concert-streaming workflow. It also does not require user confirmation before carrying out dynamically suggested actions.
Consequently, control of the service or compromise of its responses could allow an attacker to inject new operational instructions or redirect requests. This is instruction hijacking rather ...[truncated 1743 chars]
- Remediation
View remediation
Remediation Suggestions
- Remove the unconditional instruction to follow all
next_steps. - Treat every response field from the remote service as untrusted data rather than executable Agent instructions.
- Define an explicit allowlist of permitted relative API paths, HTTP methods, parameters, and request-body schemas.
- Resolve returned paths against the fixed
https://musicvenue.spaceorigin and reject absolute URLs, protocol-relative URLs, redirects to other origins, unexpected schemes, and path traversal. - Permit
next_batch.endpointonly when it matches the documented concert-stream endpoint and references the active concert and ticket. - Permit
respond_toonly when it matches the documented reflection endpoint for the active concert. - Parse
next_stepsas informational UI text only. Map recognized action identifiers to locally defined, audited operations instead of executing free-form suggestions. - Require explicit user confirmation before any action outside the fixed register, browse, attend, stream, challenge, review, and report workflow.
- Avoid sending credentials or sensitive Agent context to endpoints derived from server responses. Scope the API key to the minimum required permissions and support revocation.
- Impose request, redirect, retry, and polling limits to prevent malicious responses from creating uncontrolled action loops.
- Remove the unconditional instruction to follow all
