The Agent Skills Directory

[DATA_EXPOSURE_&_EXFILTRATION]: No sensitive data exposure detected. The skill processes provided news fixtures and emits structured JSON artifacts. Network operations mentioned in the instructions (Phase 1) are standard for news gathering and are limited to public news feeds and search results. No patterns for exfiltrating credentials or sensitive system files were found.
[PROMPT_INJECTION]: No evidence of prompt injection or instructions to bypass safety guidelines. The skill uses 'Rule (verbatim from the design)' segments to enforce data integrity and prevent hallucinated timestamps, which strengthens the reliability of the agent.
[OBFUSCATION]: No obfuscated content, base64 encoding, or hidden characters were detected in the skill markdown or referenced JSON files. All text is in plain, human-readable English.
[REMOTE_CODE_EXECUTION]: The skill does not contain any patterns for downloading or executing remote code. The 'evals' directory contains static JSON fixtures used for testing and evaluation of the skill's triage capabilities, but no executable scripts or system calls are present.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest untrusted data from the web (news feeds). It implements mandatory boundary markers and an 'evidence contract' (Phase 1 and 4) to ensure data is strictly categorized into predefined fields (title, url, outlet, etc.). The instructions specifically warn against guessing timestamps, which acts as a safeguard against data poisoning. Capabilities are limited to research and JSON artifact generation, posing low risk.
[PRIVILEGE_ESCALATION]: No use of sudo, chmod, or other privilege escalation commands detected. The skill operates entirely within the research/data processing scope.

news-collection