air-adk-trust

EU AI Act compliance for Google Agent Development Kit (ADK) agents.

Drop-in plugin that adds tamper-evident audit logging, PII detection, prompt injection scanning, and tool policy enforcement to any ADK agent — including multi-agent hierarchies.

Part of the AIR Blackbox trust layer ecosystem.

Quickstart

from air_adk_trust import AIRBlackboxPlugin
from google.adk import Agent

plugin = AIRBlackboxPlugin()
agent = Agent(name="my_agent", model="gemini-2.0-flash", plugins=[plugin])

That's it. Every agent action is now logged to a tamper-evident HMAC-SHA256 audit chain.

Install

pip install air-adk-trust

What It Does

The plugin hooks into all 6 ADK callback points:

Callback	What AIR Does
before_agent	Start audit record, check risk tier
after_agent	Finalize record, seal HMAC chain
before_model	Log prompt, scan PII, detect injection
after_model	Log response, scan output for PII
before_tool	Classify tool risk, enforce policy, check blocked list
after_tool	Log result, scan for PII leakage

EU AI Act Coverage

Article	Requirement	How AIR Covers It
Art. 9	Risk Management	Tool risk classification + configurable risk tiers
Art. 10	Data Governance	PII detection + optional blocking/redaction
Art. 11	Technical Documentation	Structured JSON audit export
Art. 12	Record Keeping	HMAC-SHA256 tamper-evident audit chain
Art. 14	Human Oversight	Blocked tool lists + confirmation requirements
Art. 15	Robustness	Prompt injection detection + loop limits + error tracking

Configuration

from air_adk_trust import AIRBlackboxPlugin, AIRConfig, RiskLevel

config = AIRConfig(
    risk_tier=RiskLevel.HIGH,       # LOW, MEDIUM, HIGH, CRITICAL
    pii_detection=True,              # Scan for emails, SSNs, credit cards, etc.
    block_pii=False,                 # Set True to block prompts with PII
    injection_detection=True,        # Scan for prompt injection attacks
    block_injections=False,          # Set True to block detected injections
    blocked_tools=["shell", "exec"], # Forbidden tool names
    max_consecutive_errors=5,        # Error circuit breaker
    max_loop_iterations=50,          # Loop detection limit
)

plugin = AIRBlackboxPlugin(config=config)

Verify the Audit Chain

# Check chain integrity
result = plugin.verify_chain()
print(result)  # {"valid": True, "total_entries": 42}

# Export for compliance reporting
audit_data = plugin.export_audit()

# Get recent events
events = plugin.get_recent_events(n=10)

Multi-Agent Support

ADK plugins fire for every sub-agent in a hierarchy. One plugin instance covers the entire agent tree:

from air_adk_trust import AIRBlackboxPlugin
from google.adk import Agent

plugin = AIRBlackboxPlugin()

researcher = Agent(name="researcher", model="gemini-2.0-flash", plugins=[plugin])
writer = Agent(name="writer", model="gemini-2.0-flash", plugins=[plugin])
coordinator = Agent(
    name="coordinator",
    model="gemini-2.0-flash",
    sub_agents=[researcher, writer],
    plugins=[plugin],
)

AIR Blackbox Ecosystem

Package	Framework	PyPI
air-langchain-trust	LangChain
air-crewai-trust	CrewAI
air-autogen-trust	AutoGen
air-openai-trust	OpenAI SDK
air-adk-trust	Google ADK	This package
air-blackbox-mcp	MCP Server

License

Apache 2.0