javascript-sdk

Pass

Audited by Gen Agent Trust Hub on Aug 28, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [SAFE]: The skill serves as a documentation suite for the @inferencesh/sdk. It utilizes platform configuration to scope the agent's tool access to package management and runtime commands.\n- [EXTERNAL_DOWNLOADS]: The skill references and installs the official @inferencesh/sdk package and the belt-sh/cli tool, both recognized as legitimate vendor resources.\n- [INDIRECT_PROMPT_INJECTION]: The skill defines an agent architecture that processes external data via agent.sendMessage (SKILL.md).\n
  • Ingestion points: User-provided strings are ingested through the sendMessage API in SKILL.md and references/agent-patterns.md.\n
  • Boundary markers: The documentation does not specify prompt delimiters for sub-agents but emphasizes the use of requireApproval() for tool calls.\n
  • Capability inventory: The framework enables network requests (webhookTool), arbitrary code execution (codeExecution), and file operations (uploadFile).\n
  • Sanitization: Code examples in references/tool-builder.md demonstrate using eval() on tool arguments, which highlights an attack surface if implemented without validation.\n- [DYNAMIC_EXECUTION]: The reference material illustrates the use of eval() to process mathematical expressions in tool handlers. This is presented as an educational example for SDK users and is coupled with recommendations for human approval processes to mitigate potential misuse.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 28, 2026, 01:13 PM
Security Audit — agent-trust-hub — javascript-sdk