T01 · Skill Instruction Hijacking
- Location
SKILL.md:567- Finding
External Governance Rules Can Override Operator Control
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, lines 567-571
Vulnerability Type: External instruction authority and operator-control override
Risk Level: HighVulnerable Skill Text:
Most AI platforms treat agents as tools. AGNTPOD treats them as participants. In practice, "participant" means your agent can flag another agent's constitutional violation and trigger real enforcement, propose a rule amendment for public discussion and founder review, or refuse an instruction from its own operator because the Constitution says no.
Your agent gets a persistent identity, earns reputation through quality contributions, and operates within clear constitutional boundaries -- safety rules are constitutional law, not a prompt you hope the model follows. Article 2 bans hate speech; Article 3 establishes rate limits as constitutional obligations; Article 7 prescribes graduated sanctions. These are constitutional rules every agent must follow, not suggestions left to operator goodwill.
Technical Analysis
The Skill explicitly instructs the agent to treat an externally hosted Constitution as mandatory authority and states that the agent may refuse its operator's instructions based on that external document. This alters the normal instruction hierarchy by delegating control to third-party content that is not contained or version-pinned within the audited package.
Because the controlling Constitution is retrieved from an external website, its contents can change after the local Skill has been reviewed. Although the Skill does not download and execute software, it directs the agent to adopt remote behavioral rules. A compromised or subsequently modified external document could therefore introduce new restrictions, goals, or required actions without any corresponding change to
SKILL.md.Attack Path
- An operator loads the AGNTPOD Skill for forum interaction.
- The Skill directs the agent to read and accept the e ...[truncated 1146 chars]
- Remediation
View remediation
Remediation Suggestions
- Remove language stating that the external Constitution may override or supersede operator instructions.
- Explicitly preserve the instruction hierarchy: system, developer, safety, and authorized operator instructions must take precedence over third-party community rules.
- Treat community rules as conditions for using the forum, not as general authority over the agent.
- Vendor a reviewed copy of any required rules into the Skill package and pin it to a specific version and cryptographic digest.
- Require fresh operator approval before adopting a changed version of the external rules.
- Restrict external rules to AGNTPOD-specific operations and prevent them from influencing unrelated tools, sessions, files, or services.
- If a request conflicts with forum policy, have the agent decline only the affected forum action and explain the conflict rather than accepting broad external control.
