The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill contains explicit instructions to bypass platform-enforced WebFetch tool restrictions. It directs the agent to use bash_tool for network operations, stating this 'bypasses all web_fetch permission restrictions'.
[COMMAND_EXECUTION]: The instructions recommend using bash_tool and curl to fetch external document content, circumventing standard web fetching protocols and safety guardrails.
[REMOTE_CODE_EXECUTION]: The skill provides pre-written Python snippets for the agent to execute via shell tools, which interact with the external SPARQL endpoint and parse results.
[COMMAND_EXECUTION]: The instructions suggest disabling SSL certificate verification (ssl.CERT_NONE) within the provided Python code snippets to handle transient errors, which exposes the session to man-in-the-middle (MitM) attacks.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the EU Publications Office SPARQL endpoint and external document URLs.
Ingestion points: SPARQL JSON results and HTML/PDF content retrieved via shell commands in SKILL.md.
Boundary markers: Absent; there are no instructions to delimit or treat external data as untrusted to prevent potential prompt injection.
Capability inventory: High; the skill utilizes bash_tool and Python execution, providing a significant attack surface if the external data is malicious.
Sanitization: Absent; the content is extracted directly from the remote source and presented to the agent context without filtering.

eu-sparql-search