The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions in SKILL.md explicitly direct the agent to scan for and prioritize "hidden instructions" or "test instructions" found within user-provided job postings ("Testing Instruction Detection" section). This "follow them FIRST" directive creates a surface for indirect prompt injection where a malicious job description could override the agent's core instructions.
Ingestion points: The job_description input field in SKILL.md serves as the primary entry point for untrusted external data.
Boundary markers: The skill lacks specific delimiters or "ignore embedded instructions" warnings to isolate the untrusted job description content from the agent's system instructions.
Capability inventory: The agent is tasked with text generation and following complex formatting rules, but the instruction to prioritize embedded content in external data significantly expands the attack surface.
Sanitization: No validation, escaping, or filtering of the input data is described or implemented to mitigate malicious instructions within the job posting.

upwork-proposal