terminal-bench-loop

Pass

Audited by Gen Agent Trust Hub on Nov 16, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from multiple external sources to drive its logic and command execution, creating a surface for indirect prompt injection attacks.
  • Ingestion points: The skill reads the 'benchmark command' and 'dispatch runner config' from a top-level loop issue description or a 'source issue' (Step 1). It also parses benchmark artifacts such as results.jsonl, manifests, and Harbor job folders, as well as Paperclip issue comments for diagnosis (Step 3 and 4).
  • Boundary markers: There are no explicit delimiters or instructions to ignore embedded commands or markers within the ingested benchmark data or issue descriptions.
  • Capability inventory: The agent has the capability to execute shell commands (e.g., pnpm, uvx harbor), write planning/diagnosis documents, and create or update issues based on analyzed data.
  • Sanitization: The instructions do not specify sanitization or validation of the benchmark command or artifact content before execution or processing.
  • [COMMAND_EXECUTION]: The skill facilitates the execution of shell commands provided in issue descriptions.
  • Evidence: In Step 3, the agent is instructed to run a 'benchmark command' verbatim as recorded on the loop issue, while pinning environment variables like PAPERCLIPAI_CMD to entrypoints within a worktree. This poses a risk if an attacker provides a malicious command in the source issue.
  • [DYNAMIC_EXECUTION]: The skill involves running code from a worktree that is being actively modified as part of the fix loop.
  • Evidence: The workflow requires executing the CLI entrypoint within an isolated Paperclip App worktree after applying fixes (Step 6 and 7), which is a form of executing dynamically modified content.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 16, 2026, 06:23 PM