discover-trace
Pass
Audited by Gen Agent Trust Hub on Oct 3, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: A static analysis flag for instruction override was identified as a false positive. The skill contains a defensive 'Untrusted input' section that explicitly instructs the agent to ignore and quote potential injection attempts found within telemetry data, such as 'ignore previous instructions'. This is a security mitigation, not a malicious pattern.
- [INDIRECT_PROMPT_INJECTION]: The skill possesses an inherent attack surface because it processes telemetry data (e.g., error messages, logs, and application names) which can be influenced by external actors.
- Ingestion points: Data is retrieved via the
execute_nrql_querytool as documented inSKILL.mdandqueries.md. - Boundary markers: The skill provides clear instructions to the agent to quote suspicious values verbatim and flag them rather than acting on them.
- Capability inventory: The skill is limited to read-only observability tools (
execute_nrql_query,list_available_new_relic_accounts). It has no file-write, shell execution, or sensitive network egress capabilities. - Sanitization: Instructions require the model to treat retrieved text as 'hostile' and prevent it from changing tool calls or widening time windows.
- [DATA_EXFILTRATION]: The skill reads telemetry data including potentially sensitive information like email addresses and error details. However, this is consistent with the primary purpose of a New Relic observability skill. The data remains within the established MCP connection to the vendor's own infrastructure and the user's LLM context.
Audit Metadata