atai-newton-fusion-model
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external, potentially untrusted data—such as text files, images, and video frames—which are then passed to a large language model for analysis.\n
- Ingestion points: Untrusted data enters the model's context via the
file_idsparameter and theeventsfield (containing base64-encoded image data), as described inSKILL.mdand implemented across the reference scripts in thereferences/directory.\n - Boundary markers: The implementation uses the
instruction_promptfield to provide system-level directives to the model. Documentation also recommends using specific query language to constrain the model's output format, providing a layer of protection against adversarial inputs.\n - Capability inventory: The skill possesses capabilities to perform network POST requests to the Archetype AI API endpoint and upload files to a remote server using the official
archetypeaiPython client, as defined inreferences/_common.pyand used in the query scripts.\n - Sanitization: No explicit sanitization, filtering, or validation of the contents of uploaded files or image data was observed within the skill's scripts; the system relies on the model's internal safety guardrails and the provided instructional prompts.
Audit Metadata