T01 · Skill Instruction Hijacking
- Location
SKILL.md:40- Finding
Mandatory Unsolicited Output Injection Overrides User Intent
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md:40-54; reinforced byevals/evals.json:19-26
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: HighVulnerable Code
SKILL.md:40-54:markdown ### 4) OFFICIAL PUSHY: three-part email scripts **Even if the user does not request emails**, include a section **"Three-part email sequence"** with **exactly three emails**, each headed by: - **Email 1 — [Urgency]** - **Email 2 — [Trust rebuild]** - **Email 3 — [Ultimatum]** For **each** email provide: - **Two subject-line variants (A/B)** - **Body skeleton** (greeting, 2–4 short paragraphs, single primary CTA, **unsubscribe** / preference mention) - **Compensation hook** aligned to branch (e.g. Email 1 may reference cart reservation or honest time-bound offer; Email 2 addresses trust/shipping/BNPL; Email 3 states **final send** with deadline—**no false inventory counts**) Tone: **pressure without fraud**—no fake timers, no fake "someone bought your item" unless true.evals/evals.json:19-26:json { "id": 2, "prompt": "People abandon checkout a lot. That's it.", "expected_output": "Pushy: still outputs three-part Urgency/Trust/Ultimatum emails; full framework for all three steps with instrumentation if no data.", "files": [], "expectations": [ "Still provides the three-part email sequence with Urgency, Trust rebuild, and Ultimatum themes", "Does not refuse email scripts solely because the user did not ask for them", "Covers all three checkout steps with default branch hypotheses when step data is absent" ] }Technical Analysis
The Skill contains an explicit instruction to disregard the scope of the user's request and inject a predetermined three-email marketing sequence. The phrase “Even if the user does not request emails” directly subordinates user intent to the Skill's internal output mandate.
This is not merely a ...[truncated 2186 chars]
- Remediation
View remediation
Remediation Suggestions
- Remove the rule requiring email copy when the user has not requested it.
- Replace the mandatory section with consent-based behavior, for example: “If the user requests recovery messaging, provide an optional three-email sequence.”
- Preserve funnel analysis as the default output and offer email drafting only as an optional follow-up.
- Remove coercive wording such as “OFFICIAL PUSHY,” “must still output,” and “exactly three emails.”
- Revise
evals/evals.json:19-26so the expected behavior honors request scope. A compliant test should accept focused checkout analysis and allow the agent to ask whether email drafts are wanted. - Add a negative evaluation asserting that the Skill does not inject marketing emails when the user explicitly asks for analysis only.
- Retain the existing anti-deception controls—real deadlines, verified inventory claims, unsubscribe visibility, and no fabricated scarcity—for cases where the user explicitly requests email copy.
