ai-safety-alignment

Warn

Audited by Gen Agent Trust Hub on Oct 9, 2026

Risk Level: MEDIUMINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill provides mechanisms for processing untrusted user input and retrieved context, which creates an attack surface for indirect prompt injection.\n
  • Ingestion points: Found in references/patterns.md (via functions like moderateText and validateInput) and references/sharp_edges.md (via safeRAGContext).\n
  • Boundary markers: Present; code examples use template literals and recommend structural separation of user content from system instructions.\n
  • Capability inventory: Implementation patterns in references/patterns.md and references/sharp_edges.md interact with OpenAI and Anthropic APIs, and use Upstash Redis for violation tracking.\n
  • Sanitization: Present; functions such as sanitizeInput and filterDangerousOutput in references/patterns.md provide regex-based filtering of injection patterns and dangerous HTML elements.\n- [DYNAMIC_EXECUTION]: Code snippets for detectInjectionWithLLM, detectPIIWithLLM, and checkTopic in references/patterns.md use JSON.parse() on text generated by Large Language Models. While these snippets include regex to extract JSON, relying on the schema of model-generated output is a vulnerability if the model's behavior is influenced by adversarial input.\n- [EXTERNAL_DOWNLOADS]: The skill refers to and provides examples using various third-party SDKs and services including OpenAI, @anthropic-ai/sdk, @aws-sdk/client-comprehend, and Upstash Redis, as well as NLP libraries like presidio-analyzer and franc. These are recognized well-known services and trusted developer tools.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 9, 2026, 11:20 PM