The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted data from LLM experiment events (inputs, outputs, and metadata) fetched via Datadog APIs. A malicious experiment record could contain instructions designed to manipulate the agent's analysis or output. The skill lacks explicit boundary markers or sanitization logic for this ingested content, although it does instruct the agent to mask PII.\n
Ingestion points: list_llmobs_experiment_events and get_llmobs_experiment_event tool calls in SKILL.md.\n
Capability inventory: File system writing, shell command execution via the pup CLI, and Datadog notebook creation.\n
Boundary markers: Not specified in the workflow instructions.\n
Sanitization: Instructions are limited to PII masking and do not cover adversarial input filtering.\n- [COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands using the pup CLI tool for its data processing backend. While experiment IDs are validated as UUIDs, other parameters such as metric labels or dimension values are derived from the experiment data itself. If these values contain shell metacharacters and the agent invokes the CLI via a shell without proper escaping, it could lead to command injection. Note that the data source is the user's own Datadog infrastructure.\n- [EXTERNAL_DOWNLOADS]: The skill provides instructions for the user to add an MCP server from https://mcp.datadoghq.com. This domain is the official platform for the trusted vendor and author (Datadog), and the reference is part of the intended setup process.

llm-obs-experiment-analyzer