eval-triage
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill fetches documentation from the vendor's official web domain (langwatch.ai) and uses the npx utility to download the langwatch package from the npm registry for reporting purposes.- [COMMAND_EXECUTION]: The agent is instructed to execute multiple subcommands of the langwatch CLI, such as experiment listing, result retrieval, and evaluator inspection, to perform diagnostic tasks.- [DATA_EXFILTRATION]: The skill includes a diagnostic submission feature via the langwatch report command, which sends session transcripts and summaries to the vendor's infrastructure. This capability is protected by mandatory user-approval flags and performs local scrubbing of sensitive patterns like API keys and PII before transmission.- [INDIRECT_PROMPT_INJECTION]: The skill operates on external experiment data and traces, which could potentially contain adversarial instructions. \n
- Ingestion points: Untrusted data is retrieved from the LangWatch platform using commands such as langwatch experiment results, langwatch dataset get, and langwatch trace get. \n
- Capability inventory: The skill has the ability to execute shell commands and transmit data via the npx langwatch report utility. \n
- Boundary markers: No explicit prompt delimiters or instructions are provided to the agent to isolate external data from its primary instructions. \n
- Sanitization: The reporting tool automatically scrubs common sensitive data types (secrets, phone numbers, emails) from the telemetry before it is sent.
Audit Metadata