The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data such as APKs, JavaScript bundles, and network captures, which constitutes an attack surface for indirect prompt injection.
Ingestion points: Artifacts (APK, JS, SQLite, WASM) and network captures (HAR, SAZ, CDP streams) are processed in various modules (e.g., FLUTTER-ARTIFACTS.md, JAVASCRIPT.md).
Boundary markers: Emphasizes using Evidence IDs and separating observations from inferences (SKILL.md Rule 6) and warns against inventing data.
Capability inventory: Utilizes subprocesses for Node.js scripts and specialized binaries (WABT) and performs file system operations for exporting and analyzing sources.
Sanitization: Includes best-effort redaction rules for tokens and secrets in generated API specifications, with instructions for manual review to prevent credential exposure (TRAFFIC-TO-OPENAPI.md).
[EXTERNAL_DOWNLOADS]: The skill mentions using npx for tool execution, which may involve downloading packages from official registries. It also references external tools and repositories (e.g., morluto/rea, browserbase/skills) as dependencies for its reverse-engineering workflows. These references are documented as part of the tool's necessary operational context.
[DATA_EXPOSURE]: Detailed guidance is provided for managing sensitive data found in network captures. The skill explicitly warns users about the limitations of automated redaction and the risk of unredacted tokens appearing in generated code or reports, instructing users to perform manual verification before sharing output.

reverse-engineering