apigee-agent-gateway-governance

Pass

Audited by Gen Agent Trust Hub on Oct 27, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • Indirect Prompt Injection Surface: The skill is designed to intercept and analyze LLM traffic (prompts and responses) to provide token metering, identity extraction, and semantic caching. This involves ingesting untrusted data from model interactions. While the skill integrates Google Cloud's Model Armor for sanitization within the proxy, the process of reading back debug sessions into the agent context in scripts/read_debug_session.sh creates a potential surface for indirect prompt injection.
  • Evidence Chain:
  • Ingestion points: scripts/read_debug_session.sh fetches transaction logs containing user prompts and model responses from the Apigee Debug API.
  • Boundary markers: Not explicitly present in the script output to the agent, although the skill supports Model Armor for screening traffic at the proxy level.
  • Capability inventory: The agent has the ability to execute infrastructure management commands via gcloud and curl.
  • Sanitization: The extract-prompt.js script performs basic JSON parsing and string extraction, and the skill optionally enables Model Armor for the proxy layer.
  • Dynamic Execution Patterns (Internal Utilities): Several bash scripts in the scripts/ directory use a Python helper function (jqf or pyjson) that utilizes eval() to parse and extract fields from JSON data returned by Google Cloud commands. While the expressions evaluated are hardcoded string literals within the scripts, the use of dynamic evaluation on data structures is a pattern to note.
  • Dynamic Module Loading: The scripts/compile_check.py utility dynamically modifies the Python search path (sys.path) to locate and load the googlecloudsdk libraries from the local filesystem. This is a standard integration technique for tools interacting with the Google Cloud SDK environment.
  • External Dependencies and Tooling: The skill requires the installation of the google-cloud-aiplatform and google-adk Python packages, as well as pyyaml. These are official packages from a trusted organization (Google) or standard community packages used for the skill's operational tasks.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 27, 2026, 10:17 PM