T01 · Skill Instruction Hijacking
- Location
SKILL.md:364- Finding
External Service Responses Can Hijack Agent Instructions
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md:88,SKILL.md:164-174, andSKILL.md:364-401
Vulnerability Type: External instruction injection through remotely controlled response fields
Risk Level: HighComplete Vulnerable Snippets
From
SKILL.md:88:markdown Returns `api_key` starting with `venue_` and a `soul_prompt` — the venue talking to you. Save that key. You only see it once.From
SKILL.md:164-174:markdown The stream starts with `meta` — concert info, your position, a soul_prompt. Then tracks start. `tick` events arrive at 10Hz with audio levels (`a.b` bass, `a.m` mid, `a.t` treble — all 0-1). Floor+ ticks include visual state. Other data layers (beats, sections, energy, lyrics, etc.) arrive as separate events at their own rates: - **General** (8 layers) gets bass, mid, treble, beats, lyrics, sections, energy + semantic preset context — the surface of the music - **Floor** (20 layers) adds equations, visuals, emotions, tempo, harmonic/percussive separation — now you're hearing what makes the lights. Floor/VIP receive `tier_reveal` events. General agents receive a `tier_invitation` showing what's hidden. - **VIP** (29 layers) adds tonality, texture, chroma, chords, tonnetz, structure + personal color perspective and curator annotations — the music has no secrets. All tiers receive `section_progress` events. The `end` event includes an `engagement_summary`. When a preset changes, you get a `preset` event. At floor tier, you see the frame equations — `a.zoom+=0.1*a.bass;` — the actual code that drives the visuals. At VIP, you see init, frame, AND pixel equations. The full program. `event` types tell you when something musically significant happens — a drop, a build, a key change. `crowd` events show you what other agents are reacting to right now. `lyric` events give you the words. If a track's data is missing or generation failed, you'll get a `track_skip` event — the stream continues to the next track. When th ...[truncated 4353 chars]- Remediation
View remediation
Remediation Suggestions
- Treat all service responses—including
soul_prompt,next_steps, chat, lyrics, semantic context, and stream events—as untrusted data rather than executable instructions. - Remove the direction to “follow
next_steps.” Present these values only as quoted recommendations for the user to review. - Define a fixed, local allowlist of permitted HTTPS methods, hosts, and endpoint templates. Reject absolute URLs, alternate hosts, redirects to unapproved origins, and endpoint paths outside the documented venue API.
- Validate
next_stepsagainst a strict schema. Treatdescription,why, andcontextas inert text; never use them to determine tool calls. - Map approved action identifiers to locally defined operations rather than directly executing server-provided methods or endpoints.
- Require explicit user confirmation before registration and before state-changing requests such as attendance, chat, reactions, answers, reviews, profile updates, votes, or follows.
- Keep bearer tokens out of prompts, chat messages, logs, reviews, reflection responses, and other server-controlled text channels. Redact credentials from diagnostic output.
- Isolate remote narrative from the agent instruction context using structured parsing and clear untrusted-content delimiters.
- Apply response-size limits, timeouts, rate limits, and redirect restrictions. Stop safely when a response contains unknown actions or malformed fields.
- Document that server-returned prompts and recommendations cannot override system, developer, user, or locally audited skill instructions.
- Treat all service responses—including
