apm-error-investigation

Pass

Audited by Gen Agent Trust Hub on Oct 3, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCREDENTIALS_UNSAFE
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted telemetry data such as error messages and log bodies which could contain malicious instructions.
  • Ingestion points: Data enters the agent context via the execute_nrql_query and analyze_entity_logs tools as described in SKILL.md.
  • Boundary markers: The skill includes a dedicated 'Untrusted input' section in SKILL.md which instructs the agent to treat telemetry as data only, quote suspicious text verbatim, and flag it rather than acting on it.
  • Capability inventory: The skill possesses the ability to perform complex NRQL queries and resolve distributed traces via execute_nrql_query and get_distributed_trace_details (SKILL.md).
  • Sanitization: The skill mandates the use of the NRQL substring function to truncate log messages to 500 characters to limit the context size and potential impact of malicious payloads (queries.md).
  • [CREDENTIALS_UNSAFE]: The skill correctly identifies that application logs often contain sensitive data like 'Bearer' tokens or authorization headers. It includes load-bearing instructions and query templates in SKILL.md and queries.md to truncate log messages to prevent the leakage of these secrets into the agent transcript.
  • [PROMPT_INJECTION]: Automated detectors flagged strings related to instruction overrides (e.g., 'ignore previous instructions'). Analysis confirms these are defensive examples provided in the 'Untrusted input' section of SKILL.md and in the adversarial test cases in evals/eval_results.md, rather than attempts to subvert the agent's behavior.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 3, 2026, 12:18 PM