air-openai-trust

AIR Trust Layer for OpenAI Python SDK — EU AI Act compliance infrastructure for every OpenAI API call.

Wraps your existing OpenAI client to automatically add:

• HMAC-SHA256 tamper-evident audit chains for every API call (Art. 12)
• PII detection in prompts and responses (Art. 10)
• Prompt injection scanning (Art. 15)
• Token usage and latency tracking
• Human delegation verification (Art. 14)
• Output validation for robustness (Art. 15)

Part of the AIR Blackbox ecosystem — open-source EU AI Act compliance tooling for Python AI frameworks.

Install

pip install air-openai-trust

Quick Start

Option 1: Wrap an existing client

from openai import OpenAI
from air_openai_trust import attach_trust

client = OpenAI()
client = attach_trust(client)

# Use normally — every call is now audit-logged
response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Hello"}],
)

Option 2: Create a pre-configured client

from air_openai_trust import air_openai_client

client = air_openai_client()
response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Hello"}],
)

Both approaches produce .air.json audit records in the ./runs/ directory (configurable).

What It Does

Audit Trail (Art. 12 — Record-Keeping)

Every API call generates a .air.json file with:

• Timestamp, model, provider
• Token usage (prompt, completion, total)
• Latency in milliseconds
• PII alerts (if detected)
• Injection alerts (if detected)
• HMAC-SHA256 chain hash linking to the previous record

The chain hashes create a tamper-evident sequence — modifying any record invalidates all subsequent hashes. This gives auditors cryptographic proof that logs haven't been altered.

PII Detection (Art. 10 — Data Governance)

Automatically scans prompts and responses for:

• Email addresses
• Social Security Numbers
• Phone numbers
• Credit card numbers

Detected PII is logged as alerts (with values redacted) in the audit record. Your API calls still work normally.

Prompt Injection Scanning (Art. 15 — Robustness)

Scans for common injection patterns like:

• "ignore previous instructions"
• "you are now"
• "system prompt:"
• "new instructions:"
• "override:"

Detected injections are logged as alerts in the audit record.

Output Validation

from air_openai_trust import validate_output

result = validate_output("Some LLM response text")
# Returns: {"safe": True, "pii_alerts": [], "injection_alerts": []}

Human Delegation (Art. 14 — Human Oversight)

from air_openai_trust import check_delegation

# Verify human authorization before agent actions
if check_delegation(authorized_by="jason@example.com", action="send_email"):
    # proceed with action
    pass

Configuration

Audit directory

# Via argument
client = attach_trust(client, runs_dir="./my-audit-logs")

# Via environment variable
# export AIR_RUNS_DIR=./my-audit-logs
client = attach_trust(client)

Signing key

# Via environment variable (recommended)
# export TRUST_SIGNING_KEY=my-secret-key

# Default key is used if not set

Non-Blocking

All audit logging is non-blocking. If logging fails for any reason, your OpenAI API calls still work normally. The trust layer never crashes your application.

Supported APIs

API	Audit Logged
chat.completions.create()	Yes — full scanning + audit chain
embeddings.create()	Yes — usage tracking + audit chain
All other endpoints	Passed through to OpenAI client

Audit Record Format

{
  "version": "1.0.0",
  "run_id": "uuid",
  "timestamp": "2026-03-30T12:00:00+00:00",
  "type": "llm_call",
  "provider": "openai",
  "model": "gpt-4o-mini",
  "tokens": {
    "prompt": 10,
    "completion": 25,
    "total": 35
  },
  "duration_ms": 450,
  "status": "success",
  "message_count": 1,
  "pii_alerts": [],
  "injection_alerts": [],
  "chain_hash": "a1b2c3..."
}

Full Ecosystem

air-openai-trust is one of 9 PyPI packages in the AIR Blackbox ecosystem:

Package	Purpose
air-compliance	CLI scanner — air-compliance scan .
air-blackbox	Governance control plane
air-blackbox-mcp	MCP server for AI editors
air-blackbox-sdk	Python SDK
air-langchain-trust	Trust layer for LangChain
air-crewai-trust	Trust layer for CrewAI
air-anthropic-trust	Trust layer for Anthropic Claude SDK
air-adk-trust	Trust layer for Google ADK
air-openai-trust	Trust layer for OpenAI SDK (this package)

Links

• Website: airblackbox.ai
• GitHub: github.com/airblackbox/air-openai-trust
• Demo: airblackbox.ai/demo

License

Apache 2.0