agent-observability-auto-experiment
Pass
Audited by Gen Agent Trust Hub on Jul 29, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [PROMPT_INJECTION]: The skill contains strong instructional language and directives to 'ignore' embedded instructions. Analysis confirms these are defensive measures intended to protect the agent and LLM-judge from malicious content within external datasets (indirect prompt injection) rather than attempts to bypass agent safety filters.
- [COMMAND_EXECUTION]: The skill generates an evaluation harness script (
eval_harness.py) and executes it locally to measure code improvements. This is a primary, documented function of the skill. - [DATA_EXFILTRATION]: The skill interacts with Datadog services using authorized MCP tools and the
pupCLI. The instructions explicitly restrict data transmission to the user's own Datadog organization and prohibit the harvesting or logging of API keys. - [SAFE]: The evaluation harness template implements a security control using zero-width spaces (U+200B) to defang potential boundary-escape attempts in XML-style tags within processed text, preventing untrusted data from overriding the evaluation rubric.
Audit Metadata