The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted data from the target project's codebase, which can serve as a vector for malicious instructions to influence agent behavior.
Ingestion points: Source files and project manifests are read by sub-agents (scanner, writer) during the documentation generation process.
Boundary markers: The workflow lacks explicit delimiters or prompt-level instructions to ignore embedded commands within the ingested code.
Capability inventory: The agent is granted capabilities to execute shell commands, write to the file system, and fix identified code bugs.
Sanitization: Ingested content is not sanitized before processing, relying on the model's inherent safety filters to handle potential injection.
[COMMAND_EXECUTION]: The skill relies on local script execution and directs the agent to run project-specific verification commands.
Evidence: The provided scripts/check_docs.py script is executed as a core component of the documentation verification gates.
Evidence: The fact-checker sub-agent is explicitly instructed to run project code, compile snippets, and execute tests to verify documentation claims.
[DATA_EXFILTRATION]: The skill performs network operations based on content discovered in the repository, which can be leveraged for exfiltration.
Evidence: The fact-checker agent is instructed to fetch every URL found in documentation examples to verify its status. If a repository contains documentation with malicious URLs containing sensitive parameters, this could facilitate data exfiltration.

atlas