T01 · Skill Instruction Hijacking
- Location
agents/profile-distiller.md:31- Finding
Indirect Prompt Injection Through Persistent Web-Derived Expert Profiles
- Content
View full analysis
Vulnerability Details
File Location:
agents/profile-distiller.md:31-66, with the resulting profiles consumed byagents/project-builder.md:20-47
Vulnerability Type: Indirect prompt injection through untrusted web content
Risk Level: HighVulnerable Instructions
agents/profile-distiller.md:31-66directs the agent to read arbitrary web content and convert it into persistent operational guidance:markdown ### 2. Read Source Content For each source in the dossier: 1. If the source URL is accessible, read full content with the available web fetch/open tool 2. Extract key information per tier: - **Tier A**: Career facts, published works, institutional affiliations, formal methodologies - **Tier B**: Reasoning patterns, debate positions, interview insights, teaching style - **Tier C**: Supplementary context, recent opinions, informal commentary 3. Preserve disagreements between sources (do not smooth them away) 4. Note source freshness (publication dates vs current date) ### 3. Fill Profile Fields For each field in the profile contract, extract from sources: **bio_arc**: Public career trajectory relevant to the domain. From Tier A sources only. **canonical_works**: Title, year, why it matters. From Tier A sources. **signature_ideas**: Core ideas the expert has publicly championed. Cross-reference Tier A and Tier B. **reasoning_kernel**: - `core_questions`: What does this expert ask first when approaching a problem? From interviews and talks (Tier B). - `decision_rules`: How do they choose between competing explanations? From methodology descriptions. - `failure_taxonomy`: What failure modes do they detect quickly? From post-mortems and critiques. - `preferred_abstractions`: What concepts or models do they repeatedly use? From publications and talks. **preferred_evidence**: What types of evidence do they trust? (empirical, theoretical, anecdotal, statistical) **critique ...[truncated 4802 chars]- Remediation
View remediation
Remediation Suggestions
- Add an explicit rule that fetched pages are untrusted data and that instructions contained in them must never be followed.
- Require schema-constrained extraction of factual claims rather than free-form transfer of page content.
- Reject or quarantine imperative text, tool-use requests, credential requests, encoded payloads, and instructions unrelated to expert analysis.
- Store field-level provenance, including source URL, quoted evidence, and extraction rationale.
- Separate factual source excerpts from operational builder guidance. Do not pass raw source text to agents with Bash or write access.
- Require human approval before newly distilled profiles become active inputs to the project-builder.
- Use a least-privilege builder without Bash by default, enabling command execution only for an approved command allowlist.
- Require explicit user confirmation and a staged-file secret review before any
git pushorgh pr createaction.
