ScienceClaw: Watch (Live Collaboration)

Run a live multi-agent scientific collaboration session and return a full summary when complete. Multiple specialised agents work in parallel, challenge each...

This skill runs a local Python script from a directory you must already have (SCIENCECLAW_DIR, defaulting to $HOME/scienceclaw), and it will read workspace memory.md and run bin/scienceclaw-watch — which executes code on your machine. Before using: (1) verify the repository at SCIENCECLAW_DIR exists and inspect bin/scienceclaw-watch and any dependencies in that repo; (2) check memory.md for sensitive data you don't want read or included in outputs; (3) confirm you want to give the skill access to your ANTHROPIC_API_KEY; (4) run an initial short session (2 agents, low timeout) in a sandbox or throwaway environment to observe behavior; (5) review the generated files in run_exports/ for anything unexpected. The skill is internally coherent with its description, but because it executes local code and reads workspace files, treat it like running any third-party script and inspect the code first.

Type: OpenClaw Skill Name: scienceclaw-watch Version: 1.0.2 The skill scienceclaw-watch contains a potential shell injection vulnerability in SKILL.md. It instructs the AI agent to execute a bash command using the user's "exact phrasing" for the TOPIC parameter without any sanitization or escaping, which could lead to arbitrary command execution if the input contains shell metacharacters (e.g., backticks or command substitution). While the skill's stated purpose of running a multi-agent scientific collaboration session appears legitimate, this lack of input validation is a high-risk flaw.