framework-benchmark

Pass

Audited by Gen Agent Trust Hub on Nov 5, 2026

Risk Level: SAFECREDENTIALS_UNSAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSOBFUSCATIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [CREDENTIALS_UNSAFE]: The skill accesses native authentication stores including ~/.claude/.credentials.json, ~/.codex/auth.json, and the macOS keychain to retrieve session tokens. These credentials are used to authenticate the benchmarking environment for the agent being tested.
  • [COMMAND_EXECUTION]: The benchmarking runner spawns multiple external processes, including git, docker, powershell.exe, and the claude/codex CLI tools, to manage isolated execution sessions and collect results.
  • [EXTERNAL_DOWNLOADS]: The framework performs remote Git fetches and clones from external repositories defined in user-supplied recipes to obtain project source code for benchmarking.
  • [OBFUSCATION]: The skill uses Base64 encoding for the transport of file deltas and to store a PowerShell script used for setting directory ACLs on Windows systems.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes arbitrary application code and documentation which may contain untrusted data. It mitigates this risk by providing explicit prompt instructions to assessor agents to treat all artifact content as data rather than instructions.
  • [DYNAMIC_EXECUTION]: The runner dynamically materializes and executes check scripts and test suites (e.g., acceptance.cjs, regression.cjs) within the benchmarked workspace to verify task correctness.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 5, 2026, 06:42 PM