json-canvas
Pass
Audited by Gen Agent Trust Hub on Sep 21, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill defines a format (JSON Canvas) that includes text and label attributes which accept Markdown and plain text. These serve as ingestion points for untrusted data that could contain malicious instructions designed to bypass agent safety protocols.
- [INDIRECT_PROMPT_INJECTION]: The format includes node types for 'file' and 'url' that take paths and links respectively. If an agent processes a canvas file from an untrusted source, these fields could be used to reference sensitive local files or malicious external domains.
- [INDIRECT_PROMPT_INJECTION]: Evidence Chain: 1. Ingestion points: JSON Canvas (.canvas) files containing text, label, file, and url fields. 2. Boundary markers: No delimiters or 'ignore embedded instructions' warnings are specified in the documentation. 3. Capability inventory: The agent is tasked with creating and editing these files, implying a level of trust in their content. 4. Sanitization: No sanitization or validation of the values within these JSON fields is defined beyond structural compliance.
Audit Metadata