runaway-stopper

Pass

Audited by Gen Agent Trust Hub on Nov 16, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes transcripts from other agent threads to identify looping or token-intensive behavior, which presents an attack surface for indirect prompt injection.
  • Ingestion points: In SKILL.md (Step 1), the skill uses tools.thread.get to fetch transcript data from multiple threads.
  • Boundary markers: The skill does not inject raw transcript text into its prompt; instead, it uses a JavaScript routine to extract specific metadata fields.
  • Capability inventory: The skill has access to Bash, file system tools (Read, Write), and thread management tools (thread.list, thread.get, subagent.stop).
  • Sanitization: Metadata is extracted into a structured JSON format and passed to an external local binary (gstack-runaway) for scanning, preventing the LLM from directly interpreting external content as instructions.
  • [DYNAMIC_EXECUTION]: The skill includes an embedded JavaScript block used within an execution tool to orchestrate complex thread enumeration, metadata filtering, and file operations.
  • [COMMAND_EXECUTION]: The skill invokes several local binaries provided by the platform vendor (garrytan), such as gstack-skill-start, gstack-skill-end, and gstack-runaway, to manage session lifecycle and perform resource scans.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 16, 2026, 08:17 PM