The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes and analyzes construction project data from external sources, which creates a surface for indirect prompt injection. * Ingestion points: The instructions.md file indicates the agent accepts data in CSV, Excel, and JSON formats, as well as direct user-provided file paths. * Boundary markers: There are no explicit instructions in the prompt to use delimiters or ignore potential instructions embedded within the safety data. * Capability inventory: The claw.json manifest defines a requirement for filesystem permissions, enabling the agent to read data from the local environment. * Sanitization: The instructions mandate that the agent should 'Validate inputs before processing', which provides a basic check but does not fully mitigate risks from maliciously crafted data content.

incident-reporting