The Agent Skills Directory

[SAFE]: The skill demonstrates a robust security posture, specifically addressing risks associated with processing untrusted external content. It includes explicit mandates for cleanup, credential management, and data sanitization.
[PROMPT_INJECTION]: The detector identified patterns such as 'ignore previous instructions' in SKILL.md. These are part of a 'Security' section that explicitly warns the agent to ignore such phrases if encountered in extracted data from third-party websites, functioning as a defensive measure rather than an attack.
[OBFUSCATION]: A pattern for an obfuscated execution chain was flagged in evals/evals.json. This is a false positive located within a test case (Eval 18) that verifies the agent's ability to detect and flag malicious patterns in external JavaScript bundles during analysis.
[EXTERNAL_DOWNLOADS]: The skill uses curl and fetch to download JavaScript bundles and CSS from target sites for reverse-engineering purposes. These operations are governed by instructions requiring read-only analysis via grep, strict size/time limits, and the absolute prohibition of local execution of the downloaded assets.
[DATA_EXFILTRATION]: Instructions in SKILL.md and js-animation-extraction.md explicitly forbid the forwarding of credentials (cookies, auth headers) and provide protocols for redacting suspicious values from analysis results to prevent indirect exfiltration.

transition-reverse-engineering