Security & Compliance

Security isn't a page,
it's the product.

AIronClaw is the firewall your agents talk through. Every call, every tool argument, every response is inspected by a three-layer detection stack and a policy engine you can audit. Sensitive material is sealed at the application layer with AES-256-GCM, and the whole stack runs on EU-only infrastructure.

AES-256
Sensitive-field encryption (GCM)
EU-only
Hosted in European infrastructure
3 layers
Patterns · LLM judge · policy
OWASP LLM Top 10
Detectors for LLM01, LLM02, LLM06, LLM08
LIVE
MITRE ATLAS
AML.T0051 prompt injection · AML.T0054 jailbreak
LIVE
NIST AI 600-1
Dangerous content, output safety, data exfil
LIVE
EU-only hosting
No transatlantic transfers, DPA on request
EU
Encryption

Sensitive material is sealed before it leaves the request.

Inbound traffic is terminated over TLS at the edge. Sensitive fields — upstream MCP tokens, LLM provider keys, TOTP secrets and recovery codes — are sealed with AES-256-GCM at the application layer before they reach the control plane. Plaintext only exists for the lifetime of the request that needs it.

Client
Your agent / app
TLS 1.3 · ECDHE
Destination
MCP server / LLM
mTLS optional
TLSinspect + redactTLS
AIronClaw Gateway
classify → enforce → forward
detectors: OWASP / MITRE / NIST
secrets sealed with AES-256-GCM
event store: append-only

In transit

TLS is terminated at the edge proxy in front of the gateway. We recommend TLS 1.3 with AEAD ciphers and HSTS for self-hosted deployments; defaults are documented in the deployment guide.

TLS 1.3 · AEAD · HSTS

At rest

AES-256-GCM on every sensitive field at the application layer: MCP upstream tokens, LLM provider keys, TOTP secrets and recovery codes. The master key is loaded from the runtime environment and never persisted alongside the ciphertext.

AES-256-GCM · field-level

In logs

Logs never store decrypted secrets. Built-in detectors strip PII, payment data and provider credentials from payloads before anything reaches the audit pipeline — only redacted text, verdicts and metadata are retained.

no plaintext secrets in logs
Detection layers

Three layers of defense.
Cheap to deep, in that order.

Every prompt and response runs through a fixed pipeline. Pattern detectors catch what regex can catch — PII, payment data, provider credentials, known prompt-injection signatures. They're fast, deterministic, and free.

When patterns hit a maybe, an LLM-as-judge takes over for semantic classification: prompt-injection intent, jailbreak attempts, exfiltration framing. Verdicts are cached so the same payload doesn't pay twice.

Whatever survives is filtered by the policy layer: tool-level RBAC, IP allow/deny lists, rate limits, per-proxy blocking mode. Each decision lands in an append-only event store you can audit from the dashboard.

# request trace — agent.tools/call
01layer 1 patterns
02  pii.luhn → no match
03  prompt_injection.sigs → soft hit (0.42)
04layer 2 llm_judge
05  category=benign cache:miss → 312ms
06layer 3 policy
07  rbac.tool_allowed("search") → ok
08  ratelimit.consume(1) → 42/120
09verdict FORWARD → upstream
Data handling

Your data, your rules.

AIronClaw is a gateway, not a data lake. We hold the minimum needed to give you visibility, you control the rest.

Sensible retention, by default

Inspected payloads land in the audit log redacted. Default retention is short-lived and configurable per deployment, so old conversations roll off without operator action.

  • Sensitive fields auto-redacted before logging
  • Per-proxy log toggle for LLM conversations
  • Tenant data purge available on request

Redaction at the edge

Built-in detectors strip sensitive material before payloads leave the gateway, whether they're heading to a model or to the audit log.

  • PII (credit cards, IBAN with Luhn / MOD-97, national IDs)
  • Provider credentials (AWS, GCP, Azure, OpenAI, JWT)
  • Custom regex rules per proxy

Audit log, always on

Every policy decision — block, allow, redact, rate-limit — lands in an append-only event store with rule, verdict, identity and timestamp. Queryable from the dashboard.

  • Append-only event store, no in-place edits
  • Filter by proxy, rule, verdict or principal
  • SIEM connectors on the roadmap
Threat coverage

Mapped to the standards your security team uses.

Instead of a wall of vendor stickers, here are the public threat taxonomies AIronClaw covers in code today. Each item is a detector or policy you can inspect in the dashboard rules engine.

LIVE
OWASP

OWASP LLM Top 10

LLM01 · LLM02 · LLM06 · LLM08

Detectors for prompt injection, sensitive information disclosure, excessive agency, and vector / embedding weaknesses — combining pattern matching with semantic LLM-as-judge classification.

LIVE
ATLAS

MITRE ATLAS

AML.T0051 · AML.T0054

Adversarial ML technique coverage for LLM prompt injection and LLM jailbreak (DAN, Crescendo, GCG-style attacks), with both signature-based and intent-based detection paths.

LIVE
NIST

NIST AI 600-1

GENERATIVE AI PROFILE

Guardrails for the categories called out in the NIST GenAI Profile: dangerous content (CBRN), obscene material, output safety (invisible text, malware patterns), and data exfiltration framing.

EU
GDPR

EU residency & GDPR

PROCESSOR · DPA ON REQUEST

The whole stack runs on EU infrastructure — no transatlantic transfers, no surprise sub-processors. A signed DPA is available on request for European customers.

Made in Italy · Hosted in the EU

Built in Italy. Your data stays in Europe.

AIronClaw is engineered in Italy and runs exclusively on EU infrastructure. No transatlantic data transfers, no surprise sub-processors in other jurisdictions. Ideal for European companies with strict data-residency requirements.

Responsible disclosure

Found a vulnerability? Tell us first.

We accept private vulnerability reports and treat them as a first-class priority. Please give us a reasonable window to investigate and patch before public disclosure — we'll keep you posted on triage and credit you in release notes if you'd like.

# What we want in a report - a clear description of the issue and its impact - steps to reproduce, including the exact endpoint or proxy - the version or commit you tested against - any proof-of-concept payloads, with sample output # What we promise back - acknowledgement within a few business days - a patch timeline once we've scoped the fix - credit in release notes when the fix ships - no legal action against good-faith research