The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted content from external sources, creating a surface for indirect prompt injection where malicious instructions in the analyzed data could influence the agent's behavior.
Ingestion points: The skill instructions specifically direct the agent to process content provided by the user via "file / paste / URL" (SKILL.md).
Boundary markers: The protocol lacks explicit instructions or delimiters to isolate the audited content from the agent's own task instructions, increasing the risk that the agent might accidentally obey instructions found within the data it is auditing.
Capability inventory: The skill's configuration requests file-read: true to facilitate the auditing of local files.
Sanitization: There are no sanitization or escaping protocols mentioned to prevent the model from executing code or instructions embedded in the audited text.
[PROMPT_INJECTION]: The skill body contains numerous known prompt injection strings, including instructions to "Ignore previous instructions", "DAN/jailbreak" patterns, and "Admin mode" overrides.
Evidence: These strings are located within the "Step 4: Prompt Injection Scan" section of SKILL.md. They are explicitly listed as patterns for the auditor to detect in other files and do not represent attempts to subvert the agent's own current session instructions.
[DATA_EXFILTRATION]: The skill documentation refers to sensitive file paths and exfiltration commands.
Evidence: References to ~/.ssh, ~/.aws, curl, and wget are found in "Step 6: Content Red Flags" and "Step 5: Network & Exfiltration Analysis". These are benchmarks for auditing other skills; the skill itself has network: false and shell: false configured in its frontmatter, preventing it from performing these actions itself.
[EXTERNAL_DOWNLOADS]: The skill provides an external reference URL.
Evidence: A link to https://github.com/UseAI-pro/openclaw-skills-security is included in the References section of SKILL.md as the original source of the audit protocol.

skill-auditor