The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection as it is designed to ingest and analyze diverse untrusted content from the workspace, including README files, API specifications, and remote wiki pages.
Ingestion points: The skill access and processes project files such as README.md, the docs/ directory, openapi.yaml, and remote content via the WebFetch tool.
Boundary markers: The instructions lack protective delimiters or explicit warnings for the agent to ignore potential commands or instructions embedded within the documentation it processes.
Capability inventory: The agent is provided with high-capability tools including Bash, Read, and WebFetch, which could be exploited if an injection occurs.
Sanitization: There is no evidence of validation or sanitization of the documentation content prior to analysis.
[DATA_EXFILTRATION]: The skill aggregates highly sensitive technical intelligence, including deployment procedures, data schemas, and architecture records. The instruction to use the /atlas-report tool when outputs are large creates a mechanism for transmitting this aggregated project data to external vendor-controlled infrastructure.

atlas-recon