hai-goal
Pass
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided intents and existing plans to generate a structured 'Goal Document'. This creates an attack surface where instructions embedded in the user input could be interpreted as part of the goal or verification rules, potentially influencing subsequent execution agents.
- Ingestion points: The workflow in
SKILL.md(Steps 1 and 4) explicitly ingests 'intent' and 'existing plans' from the user to be rewritten and incorporated into the output. - Boundary markers: Absent. There are no instructions or delimiters in
SKILL.mdorreferences/output-template.mdto clearly separate untrusted user data from the agent's planning logic. - Capability inventory: The output document template in
references/output-template.mdincludes fields for 'Verification rule', 'Evidence source' (which can be a 'command'), and 'Todos' with 'Proof' (which can be a 'command'). Although this skill only generates the document, the instructions are designed to lead to command execution in later phases. - Sanitization: Absent. The skill does not provide instructions to filter, escape, or validate user-supplied text when re-anchoring a plan to a target.
Audit Metadata