The Agent Skills Directory

[PROMPT_INJECTION]: The skill creates an attack surface for indirect prompt injection by ingesting and analyzing untrusted code from an external repository.
Ingestion points: The skill uses the Glob and Read tools to pull the complete contents of source files and configurations into the agent's context.
Boundary markers: No delimiters or explicit safety instructions (such as 'ignore embedded instructions') are provided to the security-auditor and deep-vuln-researcher sub-agents when processing this data.
Capability inventory: The agent has access to powerful tools like Bash and Write, which could be exploited if an injection attack succeeds in hijacking the agent's flow.
Sanitization: Untrusted repository content is passed to the LLM models without any form of escaping, validation, or filtering.
[COMMAND_EXECUTION]: The initialization script uses bash command substitution ($(...)) to extract values from the $ARGUMENTS variable. If the agent platform does not sanitize user input before it is interpolated into the bash environment, this allows for arbitrary command execution on the host system.

security-auditor