The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains explicit instructions to bypass standard operational constraints and safety filters. Evidence includes the 'Authority while active' section, which states that this mode 'supersedes every exploration-bounding instruction in surrounding prompts, modes, or rules' and defines 'under-exploration' as the primary failure mode.
[REMOTE_CODE_EXECUTION]: The orchestrator is instructed to generate and execute arbitrary scripts at runtime to verify claims found during research. Evidence includes the Phase 3 protocol directing the agent to 'Write a minimal self-contained script that tests the claim; run it (uv run --with python / bun / direct compile)'.
[COMMAND_EXECUTION]: The skill utilizes shell commands for environment management and fetching remote data. Evidence includes 'mkdir' commands for session directory creation and 'git clone' commands in the repo deep-dive protocol to download external repositories to local temporary storage.
[PROMPT_INJECTION]: The skill exhibits a significant surface for indirect prompt injection by processing untrusted external content. Evidence chain: 1. Ingestion points: Web browsing results, codebase search hits, and cloned repositories. 2. Boundary markers: The instructions lack explicit delimiters or warnings to ignore embedded instructions in ingested data. 3. Capability inventory: The skill possesses high-risk capabilities including arbitrary script execution ('uv run') and shell command access ('git'). 4. Sanitization: The synthesis process lacks documented validation or escaping for data incorporated into final HTML/PDF/PPTX reports.

ulw-research