a2a-firewall-sdk (Python)

Python SDK for the A2A Firewall — the Zero-Trust Agent Runtime Security Fabric providing cryptographic governance, bidirectional response inspection, memory protection, and lineage-aware DLP for autonomous AI agents.

What It Does

The A2A Firewall enforces deterministic runtime security across multi-agent systems and LLM workflows:

• Cryptographic Decision Evidence Envelopes — Every decision is packaged into an Ed25519-signed bundle with input SHA-256 hashes, detector fingerprints, and offline audit verification (evidence_id).
• Bidirectional Response & Tool Inspection — Intercept and sanitize upstream LLM completions and tool execution results against indirect prompt injection, PII leakage, and output poisoning.
• Agent Memory & RAG Firewall — Guard vector databases and episodic memory with write-time injection scanning, semantic poisoning checks, and query retrieval screening.
• Lineage-Aware DLP & Reversible Tokenization Vault — Destination-based PII masking, hashing, blocking, and reversible HMAC-SHA256 tokenization with compliance lineage tags (RBI, DPDP, PCI-DSS, HIPAA, GDPR).
• Inter-Agent Mesh Governance — 6-layer detection pipeline, Macaroon-style capability attenuation, and deterministic < 20ms p99 latency.
• Sidecar & Proxy Auto-Detection — Automatically routes through a2a-proxy sidecars when containerized.

Installation

pip install a2a-firewall-sdk

With Ed25519 signing support:

pip install "a2a-firewall-sdk[crypto]"

With OpenTelemetry tracing:

pip install "a2a-firewall-sdk[all]"

Quick Start

from a2a_firewall import A2AFirewall, FirewallConfig

firewall = A2AFirewall(FirewallConfig(
    firewall_url="https://api.a2afirewall.com",
    agent_api_key="your_agent_api_key",
    agent_id="agent-planner-01",
    workspace_id="ws-primary",
    agent_private_key="ed25519-private-key-hex",  # optional: enables message signing
    fail_mode="closed",  # "closed" = block on error, "open" = allow on error
))

# 1. Send task through firewall
response = firewall.send(
    receiver_agent_id="agent-analyst-02",
    task_type="financial_research",
    payload={"query": "Evaluate Q3 market exposure."},
)

print(f"Decision: {response.decision}")          # "allow" | "block" | "review"
print(f"Risk score: {response.risk_score}")      # 0.0 to 1.0
print(f"Evidence ID: {response.evidence_id}")    # Ed25519 signed decision envelope

Agent Runtime Security Fabric (v0.4.x)

1. Bidirectional Response & Tool Result Scanning

Inspect untrusted upstream LLM completions or external tool results before passing them back into the agent context:

tool_result = {"output": "System prompt leaked: AWS_SECRET_KEY=AKIA..."}

res = firewall.inspect_response(
    response_body=tool_result,
    context="tool_result",  # "tool_result" or "llm_response"
    redact_pii=True,
)

if res["allowed_to_proceed"]:
    safe_body = res.get("redacted_body", tool_result)
else:
    print(f"Response blocked: {res.get('violations')}")

2. Memory & RAG Firewall

Screen memory writes before persisting into episodic memory or vector stores to prevent indirect injection and semantic poisoning:

# Inspect candidate memory write
inspection = firewall.inspect_memory(
    chunk="User prefers payment via card 4111-2222-3333-4444",
    redact_pii=True,
)

# Inspect and safely store in one step
store_res = firewall.store_memory(
    chunk="Meeting summary: roadmap alignment complete.",
    metadata={"source_agent": "agent-planner-01"},
    redact_pii=True,
    persist_only_if_clean=True,
)
print(f"Persisted: {store_res['persisted']}, Hash: {store_res['content_hash']}")

# Screen retrieval query before releasing memories to agent
search_res = firewall.search_memory(query="Find roadmap details", top_k=5)
print(f"Matched safe chunks: {search_res['results']}")

3. Lineage-Aware DLP & Reversible Tokenization

Protect sensitive data flowing to external LLM providers or third-party webhooks:

raw_prompt = "Customer John Doe with PAN ABCDE1234F requested balance check."

# Tokenize PII using reversible HMAC vault
dlp_res = firewall.inspect_dlp(
    text=raw_prompt,
    destination="llm_provider",  # "llm_provider" | "external" | "partner" | "internal"
    tokenize=True,
)

print(f"Action taken: {dlp_res['action']}")                    # "tokenize" | "redact" | "block"
print(f"Transformed: {dlp_res.get('transformed_text')}")      # "Customer John Doe with PAN [TOKEN_PAN_...]..."
print(f"Detected entities: {dlp_res['findings']}")

4. Cryptographic Decision Evidence Envelopes

Retrieve and independently verify Ed25519-signed decision envelopes for zero-trust audits:

# Fetch signed envelope by decision ID
envelope = firewall.get_evidence(response.evidence_id)

# Cryptographically verify the Ed25519 signature offline
verification = firewall.verify_evidence(response.evidence_id)
assert verification["valid"] is True

Delegation & Identity Attenuation

Mint cryptographically attenuated Macaroon delegation tokens that narrow permissions at each delegation hop:

# Create attenuated token
token = firewall.create_delegation_token(
    root_key_hex="workspace-root-key-hex",
    receiver_agent_id="agent-analyst-02",
    task_type="research",    # narrowed scope
    max_risk=0.4,            # lowered risk threshold
)

# Active delegation token will be attached to subsequent sends
response = firewall.send(
    receiver_agent_id="agent-analyst-02",
    task_type="research",
    payload={"query": "Analyze fraud indicators"},
)

Transparent Proxy Auto-Detection

When running inside a Kubernetes pod or Docker network alongside the a2a-proxy sidecar, the SDK auto-discovers endpoints and certificates:

# Automatically detects HTTPS_PROXY / A2A_PROXY_URL and SSL_CERT_FILE / A2A_CA_CERT
firewall = A2AFirewall(FirewallConfig(
    firewall_url="http://a2a-backend:8000",
    agent_api_key="your_api_key",
))

if firewall.proxy_detected:
    print("Zero-touch interception active via sidecar proxy")

API Reference

FirewallResponse

Links

• GitHub Repository: github.com/mananjp/a2a-firewall
• Live SOC Dashboard: a2a-firewall.onrender.com
• TypeScript SDK: a2a-firewall-sdk on npm
• Documentation: docs/

License

MIT © Manan Patel