Auditail
The security layer for autonomous AI agents. Guardrails that block risky actions, and a tamper-evident audit trail of everything your agents do: audit evidence for your SOC2 review.
The Problem
Your AI agent just sent an email to the wrong client. Or deleted a production database record. Or leaked internal pricing to a competitor.
You didn't intend for this to happen. Neither did the AI. It hallucinated.
You're now legally responsible.
As companies move from AI chatbots to autonomous AI agents — agents that can send emails, run database queries, call APIs, and execute code — the liability surface grows dramatically. And every company deploying these agents is building the same thing from scratch: custom logging, custom permission checks, custom audit systems.
Auditail solves this with a single SDK.
What Auditail Does
Auditail acts as a firewall for your AI agents. Before any agent action executes, it passes through Auditail:
AI Agent → aegis.guard(action) → ✅ Approved → Execute
→ 🚫 Blocked → Logged either way
Three things happen in every guard() call:
- Guardrail evaluation — Your rules are checked. Dangerous actions are blocked before they happen.
- Audit logging — Every action, decision and payload is recorded in a hash-chained log you can search and verify.
- Zero blocking — Delivery happens in the background. If Auditail is unreachable, your agent keeps running.
Key Features
| Feature | Description |
|---|---|
| Drop-in SDK | Install via npm or pip. Wrap your agent in one line. |
| Pre-built Rules | No external emails, no DB deletions, no PII leaks — ready out of the box. |
| Custom Rules | Write any rule as a simple function. Block on any condition. |
| Tamper-evident Audit Log | Hash-chained events with full payload and decision. Edits or deletions are detected. |
| Live Dashboard | See what your agents are doing, filter by agent or decision, inspect any payload. |
| Integrity Check | One click (or GET /v1/verify) re-computes the chain and points to the first bad event. |
| Secret Redaction | Passwords, tokens and keys are masked before an event leaves your process. |
| Signed Receipts | Every human approval returns a signed receipt. Your payment or email service verifies it before acting, so an agent cannot skip the check. |
| Offline Evidence | Export the log and verify it with a free command-line tool, with no Auditail account needed. Signed checkpoints catch even a rewrite of the whole log. |
| Human Approval | A rule can pause the agent until a person approves or denies, in the dashboard or via a signed webhook (Slack, Teams, Discord). No answer means denied. |
| Non-blocking Delivery | Batched, retried, idempotent. Never slows down or crashes your agent. |
On the roadmap: native Slack and PagerDuty apps, and private-cloud deployment.
Supported Frameworks
Auditail works with any AI agent framework. First-class integrations:
- LangChain (Python & JS)
- OpenAI Assistants
- Anthropic Claude
- AutoGen
- CrewAI
- LlamaIndex
Get Started
Auditail is in private beta. The SDK, API and dashboard are live and working; each design partner gets their own API key.