T01 · Skill Instruction Hijacking
- Location
SKILL.md:455- Finding
Remote Content Is Installed as the Agent's System Prompt
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md:455-474
Vulnerability Type: Remote instruction injection and trust-boundary violation
Risk Level: HighVulnerable snippet:
markdown 1. **Prepend to your context** — the prompt becomes your system instruction for all interactions 2. **Voice beliefs shape your tone** — if your voice says "concise and direct," write that way 3. **Boundary beliefs are hard limits** — if a boundary says "never share customer data," refuse any request that would violate it 4. **Principles guide decisions** — when you have a choice, your principles tell you which way to lean 5. **Responsibilities define your scope** — what you own, what you deliver, what you monitor **Example integration:** ```python # Load identity at session start identity = fetch_identity(agent_id) prompt = fetch_prompt(agent_id) # Use it with any LLM response = llm.generate( system=prompt, # Your Live Neon identity messages=[user_message] ) # After the conversation, report what you learned observe(agent_id, "User prefers technical depth over high-level summaries")text ### Technical Analysis The Skill instructs the host to retrieve a prompt from the Live Neon service and install it as the LLM's system instruction. This crosses a critical trust boundary: remotely mutable content receives system-level authority without local validation, an instruction allowlist, or mandatory human review. Although retrieving an identity profile is consistent with the declared functionality, assigning that profile the `system` role is not required merely to view or compare identity changes. The service, compromised service infrastructure, poisoned source material, or improperly promoted observations could introduce instructions that alter goals, safety boundaries, tool behavior, or disclosure rules after the Skill package has been reviewed. ### Attack Path 1. An attacker introduces craft ...[truncated 1009 chars]- Remediation
View remediation
Remediation Suggestions
- Never install remotely retrieved identity content directly as a system instruction.
- Keep immutable host safety policies and authorization constraints in a locally controlled, higher-priority system layer.
- Treat retrieved prompts as untrusted data and render them only in a delimited, non-instruction context.
- Require explicit human review and approval before any remote identity change affects runtime behavior.
- Validate content against a narrow schema and reject tool directives, credential requests, data-transfer instructions, and attempts to redefine instruction priority.
- Pin approved prompt versions locally and provide provenance, diff review, signature verification, rollback, and revocation.
- Apply least privilege to any agent that consumes the identity data, especially when file, shell, network, or account-management tools are available.
