agent-observability-experiment-analyzer
Pass
Audited by Gen Agent Trust Hub on Jun 19, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes shell commands using the
bashtool to interact with thepupCLI (pup llm-obs ...). This is a core feature used to fetch experiment data when MCP tools are unavailable. - [EXTERNAL_DOWNLOADS]: The skill references and directs the agent to interact with Datadog's infrastructure, specifically the official MCP server at
mcp.datadoghq.comand variousapp.datadoghq.comlinks for UI navigation. - [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it ingests untrusted data from LLM experiment traces (inputs, outputs, and metadata) via the
get_llmobs_experiment_eventandlist_llmobs_experiment_eventstools. Malicious content within these traces could attempt to influence the agent's analysis. - Ingestion points: Experiment event details and summaries retrieved from Datadog (SKILL.md).
- Boundary markers: No specific delimiters or "ignore instructions" markers are used when the agent processes the content of retrieved events.
- Capability inventory: The agent can execute bash commands (
pup), write files, and create Datadog notebooks via MCP tools (SKILL.md). - Sanitization: The skill includes instructions to mask or redact PII from all user-visible outputs.
Audit Metadata