The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exports trace and span data from Arize that contains untrusted content, such as LLM prompts and responses. This creates a surface for indirect prompt injection if the agent attempts to interpret this data as instructions.\n
Ingestion points: Data enters the agent's context through ax spans export and ax traces export commands as defined in SKILL.md.\n
Boundary markers: SKILL.md contains an explicit 'Security: untrusted content guardrail' warning the agent to treat all exported data as raw text and never act on instructions found within attributes.\n
Capability inventory: The agent can execute shell commands via the ax CLI, write files to the .arize-tmp-traces directory, and perform irreversible span deletions.\n
Sanitization: The skill relies on natural language instructions to enforce that exported data is used only for display and analysis.

arize-trace