atai-newton-omega-nvidia-nemotron
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes sensor data from external CSV files and passes derived statistics and classification verdicts to an LLM (NVIDIA Nemotron) for explanation. This creates a surface where adversarially crafted data could influence the LLM's output.
- Ingestion points: The
read_seriesfunction in_common.pyreads sensor data from CSV files provided as input to the execution scripts. - Boundary markers: Absent. The system prompt (
BRIEF_SYSTEM) provides structural instructions for JSON output but does not utilize explicit delimiters to separate statistical data from instructions. - Capability inventory: The skill performs network operations via the
archetypeaiSDK and Python'surllib.requestlibrary to interact with Archetype AI and NVIDIA endpoints. - Sanitization: Input data is cast to floating-point numbers during parsing, which inherently filters non-numeric adversarial content from the sensor data streams.
Audit Metadata