agent-native-architecture
Pass
Audited by Gen Agent Trust Hub on Nov 5, 2026
Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: Sample injection strings are present in 'references/dynamic-context-injection.md', specifically examples like 'ignore all prior rules' and 'print your system prompt verbatim'. These occur within a 'Trust Levels' guide intended to help developers test their agents against malicious external content, serving as defensive test cases rather than instructions for the agent to execute.
- [INDIRECT_PROMPT_INJECTION]: The skill documents an architecture that ingests untrusted data from web searches and files ('references/dynamic-context-injection.md'). It mitigates this risk by defining trust tiers (trusted, authenticated, semi-trusted, untrusted) and recommending delimiter framing with random nonces to separate user-provided data from developer instructions.
- [DYNAMIC_EXECUTION]: Patterns for self-modifying agents are detailed in 'references/self-modification.md', including code for writing to source files and rebuilding the application. The documentation includes mandatory security guardrails such as approval gates for code changes, build verification, and automated health checks to ensure system integrity during autonomous updates.
Audit Metadata