llm-to-bedrock

Pass

Audited by Gen Agent Trust Hub on Oct 13, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSCREDENTIALS_UNSAFEINDIRECT_PROMPT_INJECTIONPRIVILEGE_ESCALATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill employs various shell commands and Python scripts for procedural automation of the migration workflow.
  • Uses git status and git rev-parse to verify the state of the repository at $REPO and ensure a clean working tree before modifications.
  • Executes bundled Python utility scripts via uv run for Bedrock service preflight checks, pricing lookups, and model baseline comparisons.
  • [EXTERNAL_DOWNLOADS]: Performs network operations to established service provider endpoints as part of its core functionality.
  • Communicates with official AWS Bedrock and Pricing APIs using boto3 to verify model access and cost estimates.
  • Sends HTTP requests to official LLM provider endpoints (OpenAI, Anthropic, Google) via urllib.request to establish quality baselines for migrated prompts.
  • [CREDENTIALS_UNSAFE]: The skill handles sensitive third-party API keys with dedicated security protocols to prevent exposure.
  • API keys are materialized directly from the environment into a local file (.source-provider-env) using restricted file permissions (chmod 600).
  • The skill automatically initializes the artifact directory with a .gitignore entry to prevent secrets from being committed to the repository.
  • Bundled scripts (source_baseline.py, resolve_source_model.py) implement a multi-stage redact function that strips key values and their variants from error messages, tracebacks, and output logs.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes external data which serves as a potential ingestion surface for malicious instructions.
  • Ingestion points: Reads and analyzes user-supplied source code and application logs from the local repository (processed by analysis and log-ingestor agents).
  • Boundary markers: Boundary delimiters or "ignore embedded instructions" warnings for untrusted data processing are absent in the top-level instructions.
  • Capability inventory: The skill can perform network operations and file writes across multiple scripts including preflight_bedrock.py and source_baseline.py.
  • Sanitization: All agent-generated outputs are validated against deterministic JSON schemas using validate_result.py before they are consumed by subsequent processing phases.
  • [PRIVILEGE_ESCALATION]: Documentation provides instructions for the user to resolve IAM permission gaps.
  • Guides users to manually update IAM policies via the aws iam put-role-policy CLI command if foundation model or inference profile access is denied.
  • These actions are presented as informative troubleshooting steps and are never executed autonomously by the AI agent.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 13, 2026, 03:30 PM