twg-bench-lite
Pass
Audited by Gen Agent Trust Hub on Aug 11, 2026
Risk Level: SAFE
Full Analysis
- [COMMAND_EXECUTION]: The skill executes vendor-provided CLI tools (twg, rovo, codex) to perform benchmarking operations, environment diagnostics, and quality judging. These actions are within the scope of the skill's primary purpose.
- [EXTERNAL_DOWNLOADS]: The skill facilitates the installation of required agent components via the twg skills install command, which is a standard management feature of the Atlassian agent platform.
- [DATA_EXFILTRATION]: The skill reads specific Atlassian configuration files (e.g., ~/.twg/auth.conf) to determine the target site for the benchmark. This access is localized to vendor-specific files and does not involve unauthorized data movement.
- [PROMPT_INJECTION]: The skill acts as an orchestrator that passes user-provided prompts to isolated sub-agents. It implements environment restrictions and instructions to maintain arm isolation and integrity during the benchmark.
- Ingestion points: User-supplied prompt input via command-line arguments or files.
- Boundary markers: Prompts are isolated within specific markdown blocks in generated sub-agent instructions.
- Capability inventory: Capabilities include local shell command execution and Atlassian MCP tool access, which are controlled via configuration policies.
- Sanitization: Prompts are handled as data within the benchmarking framework, with structured JSON reporting used for results.
Audit Metadata