bug-evidence-protocol
Pass
Audited by Gen Agent Trust Hub on Jul 23, 2026
Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill facilitates the execution of arbitrary shell commands through the 'capture_command.py' script. This is the intended functional behavior used to obtain auditable proof of bug reproduction and verification.
- [DATA_EXFILTRATION]: The skill records command outputs (stdout/stderr) and environment metadata into local JSON files. It includes a proactive redaction mechanism that uses regular expressions to mask common credential formats, including Bearer tokens, API keys, and platform-specific tokens (GitHub/NPM), reducing the risk of accidental sensitive data exposure.
- [PROMPT_INJECTION]: The 'SKILL.md' instructions enforce strict authority boundaries, explicitly requiring the agent to obtain user approval before performing destructive writes, configuration changes, or crossing privacy and security boundaries.
- [PROMPT_INJECTION]: There is a potential attack surface for indirect prompt injection when the agent processes command outputs containing untrusted data. 1. Ingestion points: Command output (stdout/stderr) captured in 'capture_command.py' and processed by 'classify_evidence.py'. 2. Boundary markers: Data is stored in structured JSON fields, and instructions mandate manual artifact inspection. 3. Capability inventory: Arbitrary shell command execution and local file access. 4. Sanitization: Built-in secret redaction and output length truncation.
Audit Metadata