The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection by design. It instructs the agent to ingest and analyze untrusted external data, including client-provided documents (T&Cs, contracts, emails) and content from live websites. This could allow an attacker to embed instructions in those sources to manipulate the agent's drafting process or benchmarks.
Ingestion points: SKILL.md (and its French counterpart) directs the agent to analyze client documents and research URLs to understand the business and identify technical elements.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are present in the instructions.
Capability inventory: The agent is tasked with document drafting and competitive benchmarking based on the ingested data.
Sanitization: The workflow does not include steps for sanitizing external text or validating that the analyzed content contains only descriptive data rather than instructions.

privacy-policy-malik-taiar