The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external PDF documents, creating a vector for indirect prompt injection attacks where malicious instructions could be embedded in the guides.
Ingestion points: Documents are downloaded from the official messervices.cyber.gouv.fr portal and converted to text via pdftotext (as described in SKILL.md).
Boundary markers: The skill includes instructions to 'locate and cite' sources exactly and avoid 'pre-digesting' the content, which provides a logical constraint but does not prevent adversarial injection.
Capability inventory: The skill uses network fetching, file writing (guide.txt), and shell command execution (pdftotext).
Sanitization: No explicit filtering or sanitization of the extracted PDF text is mentioned.
[COMMAND_EXECUTION]: The workflow involves the execution of the pdftotext command-line utility to convert downloaded PDF files into text format for analysis.
Evidence: The SKILL.md file explicitly instructs the agent to use pdftotext -layout guide.pdf guide.txt.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to fetch a catalogue and security guides from official sources.
Evidence: The skill fetches data from messervices.cyber.gouv.fr (official French government portal). This is a well-known and trusted source for cybersecurity documentation.

anssi-guides