The Agent Skills Directory

[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill instructions in SKILL.md mandate the use of npx -y rea-agents@latest doctor and npx -y rea-agents@latest setup for environment alignment and provider troubleshooting. The use of npx with the -y flag automatically installs and executes the specified package from the NPM registry without manual confirmation, while the latest tag prevents version pinning, creating a vector for unverified remote code execution.
[EXTERNAL_DOWNLOADS]: The skill's setup process involves downloading external software, including the 'Hopper' disassembler, as noted in the readiness instructions.
[DYNAMIC_EXECUTION]: The run_controlled_replay capability described in references/controlled-replay.md allows for the dynamic execution of extracted JavaScript modules. Although the skill specifies this must occur within a sandbox and requires explicit operator approval, it remains a high-risk capability when processing untrusted code.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest untrusted third-party data such as native binaries, mobile application packages (APK, IPA), and browser runtime memory. This ingestion occurs in a context where the agent has tools for filesystem writes and shell command execution.
Ingestion points: Native binaries, extracted JavaScript trees, and browser targets.
Boundary markers: Instructions require approved: true flags and use of specific plan digests before execution.
Capability inventory: Includes filesystem extraction (extract_artifact), dynamic replay (run_controlled_replay), and shell-based setup tools (npx).
Sanitization: Employs sandboxing for module execution and verification specs to validate reconstructed code.

reverse-engineer-anything