work-tournament

Warn

Audited by Gen Agent Trust Hub on Jun 14, 2026

Risk Level: MEDIUMCOMMAND_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill implements an "auto-discovery" mechanism for repository commands. It identifies and executes scripts defined in package.json (e.g., build, typecheck, test, lint) and Makefile targets.
  • Evidence: Step 3.5 instructions state: "Auto-discover the oracle commands from the repo: package.json scripts (build, typecheck/tsc, test, lint), Makefile targets... In each worktree run, in order: build → typecheck → test → lint."
  • Risk: If the repository is untrusted or has been compromised, an attacker can define malicious shell commands in these configuration files which the agent will then execute with its current privileges.
  • [COMMAND_EXECUTION]: The skill utilizes git worktrees and subprocesses to manage parallel implementations, involving multiple shell operations.
  • Evidence: The skill uses git worktree commands and background execution for sub-agents across multiple steps.
  • [DATA_EXFILTRATION]: The skill reads project source files and writes execution journals to the local filesystem.
  • Evidence: "Read every changed file in the worktree... write quality-journals/.json (under your agent's data dir)."
  • Risk: No external network exfiltration patterns (like curl or wget to unknown domains) were detected; data remains within the local environment and designated log directories.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jun 14, 2026, 09:42 PM
Security Audit — agent-trust-hub — work-tournament