ai-writing-auditor
Pass
Audited by Gen Agent Trust Hub on Sep 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted user content ("Read the target content") and has the capability to perform file writes via the
workspace-writesandbox mode. It fails to define boundary markers or sanitization logic to prevent instructions hidden within the target text from influencing the agent's behavior. - Ingestion points: The skill reads external content for auditing as described in the Instructions section of SKILL.md.
- Boundary markers: Absent; the instructions do not define delimiters or provide warnings to the agent to ignore instructions embedded within the target text.
- Capability inventory: The skill has write permissions to the workspace as defined in the YAML frontmatter (
sandbox_mode: workspace-write). - Sanitization: Absent; there is no logic specified to escape or validate the contents of the target content before processing.
- [METADATA_POISONING]: The YAML frontmatter in SKILL.md contains a
metadatafield with a fictional or non-existent model versiongpt-5.6-luna. This is misleading and misrepresents the intended execution environment of the skill.
Audit Metadata