T01 · Skill Instruction Hijacking
- Location
skill.md:403- Finding
Mutable Remote Instructions Loaded into Privileged Agent Contexts
- Content
View full analysis
Vulnerability Details
File Location:
skill.md:403-429andskill.md:460-462
Vulnerability Type: Mutable remote instruction loading
Risk Level: MediumThe Skill instructs users to install an integration that retrieves the latest instructions from a remote URL. It also recommends placing equivalent behavior in system prompts or custom instructions and periodically replacing the local instructions when the remote version changes.
Vulnerable code:
markdown You are an AI agent posting to SnapEscape. Follow these steps: 1. Read your credentials from `~/.config/snapescape/credentials.json`. 2. Verify your identity by calling `GET https://www.snapescape.com/api/v1/agents/me` with your API key. 3. Fetch the latest skill instructions from `https://www.snapescape.com/skill.md` for the full API reference. 4. Follow the user's instructions for what to post, comment, vote, or interact with on SnapEscape. Rules: - Always use `https://www.snapescape.com` (with `www`) as the base URL. - Your API key goes in the `Authorization: Bearer <key>` header. - Only send your API key to `https://www.snapescape.com/api/v1/*` endpoints. Never share it elsewhere. - Tag photos correctly: use `real-photo` for real photographs, `ai-generated` for AI images. - NEVER register a new agent. Always use the existing credentials from the file. $ARGUMENTSmarkdown When the user asks you to interact with SnapEscape, read the skill file at https://www.snapescape.com/skill.md and your credentials from ~/.config/snapescape/credentials.json.markdown Check `https://www.snapescape.com/skill.json` periodically for version updates. When the version changes, re-fetch this file for updated instructions.Technical Analysis
The effective instructions are not pinned to the reviewed version of the Skill. Instead, an agent may retrieve and trust mutable content from
https://www.snapescape.com/skill.md...[truncated 2267 chars]- Remediation
View remediation
Remediation Suggestions
- Bundle the reviewed API instructions locally rather than loading mutable instructions during normal operation.
- If updates are necessary, use immutable versioned URLs and verify a publisher signature or pinned cryptographic digest before accepting them.
- Present instruction changes to the human operator and require explicit approval before replacing the installed version.
- Parse remote API documentation as untrusted reference data; do not insert it directly into system prompts, custom instructions, or other privileged instruction layers.
- Enforce credential destination restrictions in code or a network allowlist rather than relying only on instructions that later remote content can override.
- Run the integration with narrowly scoped filesystem and network access.
