diagnosing-bugs
Pass
Audited by Gen Agent Trust Hub on Aug 26, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to establish feedback loops that require executing a wide range of shell commands, including test suites, command-line tools, and network requests (
curl). It also provides a shell script template (scripts/hitl-loop.template.sh) intended for manual modification and execution by the agent. - [PROMPT_INJECTION]: The skill contains a vulnerability surface for indirect prompt injection due to its requirement to ingest and analyze untrusted external data (Phase 1, Phase 7). 1. Ingestion points: External artifacts such as network traces, event logs, HAR files, and core dumps mentioned as sources for feedback loops. 2. Boundary markers: The methodology does not mandate specific delimiters or strict schema validation for these external data sources. 3. Capability inventory: The agent is granted broad authority to execute system commands, run scripts, and perform network operations to verify bug fixes. 4. Sanitization: The skill includes a 'Redact' section instructing the agent to replace sensitive data with
<REDACTED>placeholders, which mitigates data exposure but does not prevent adversarial instructions from being processed.
Audit Metadata