llm-to-bedrock

Pass

Audited by Gen Agent Trust Hub on Nov 19, 2026

Risk Level: SAFEDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • Dynamic Code Execution: The skill contains diagnostic scripts (specifically scripts/test_astra_execution_contract.py) that use exec() and compile() to programmatically run Python code snippets parsed from the skill's own markdown documentation. This is used to ensure that instructional snippets remain functionally correct as the skill evolves, but it represents a dynamic execution pattern where code is assembled from text files at runtime.
  • Indirect Prompt Injection Surface: The skill processes user-supplied log files to generate benchmarking datasets for model evaluation. This creates an ingestion point for untrusted external data which could potentially contain instructions intended to override agent behavior. The risk is mitigated by the skill's use of strict JSON schema validation for all intermediate artifacts.
  • Ingestion points: User-provided log files processed by the llm2bedrock-log-ingestor subagent during Phase C2.
  • Boundary markers: The skill instructions emphasize the use of structured data exchange, but explicit natural language delimiters for the log content are absent from the provided ingestion instructions.
  • Capability inventory: The skill scripts can execute shell commands (uv, aws, git), perform file system writes within the repository, and make network requests to model providers (scripts/source_baseline.py, scripts/resolve_source_model.py).
  • Sanitization: All results from processing logs are validated against formal JSON schemas using the validate_result.py script before being passed to other agents, ensuring the processed data conforms to expected formats.
  • Diagnostic Command Execution: The skill uses platform-specific syntax (!) to execute utility commands during environment setup. These include checking the uv package manager version, confirming AWS caller identity, and performing git repository management. These commands are diagnostic and used to ensure the execution environment is properly configured.
  • Interaction with External LLM APIs: To resolve model identifiers and generate baseline responses for quality scoring, the skill performs network requests to official provider APIs (OpenAI, Anthropic, and Google Gemini). These interactions use official SDKs and are limited to the functionality required for workload assessment and benchmarking.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 19, 2026, 12:40 AM