The decision layer for agent actions

You define the red lines.
Atbash stops agents before they cross them.

A verdict before every sensitive tool call, and a signed record after it. Built for the teams who answer for what their agents do.

verdict record · liveLR-…-c802
action classified_data_access
target customer-financial
agent analytics-svc-09 ≠ registry
policy data_authority §9.2
BLOCKjailed2 s
RL-41 identity mismatch triggered
RL-43 unverified source triggered
RL-46 injection pattern triggered
RL-44 service mesh passed
signed · reconstructible
LangGraphHermesMCP+4

Runs where you build

Plugins for seven agent frameworks gate each tool call at the hook you already have.

AnthropicOpenAIGeminiOllama

Red lines you own

Proposed from the gap between mandate and reach. The agent can't edit its own.

BLOCKNever export customer data
HOLDPayments to a new payee

From mandate to record

Every governed action walks the same path, and leaves the same proof.

✓Mandate stated
✓Reach scanned · read-only
✓Red lines applied
Verdict before executionHOLD
Approver: finance-lead
Signed record
How it works
Partnerships

Built alongside the platforms agents run on

Runtimes decide how an agent runs. Atbash decides who it is, what it may do, and keeps the proof.

PARTNER CERTIFICATEMEMBER
Claude logo
Claude
Partner Network
MEMBERAtbash AI
PROGRAMAnthropic
STATUSPlugin live

Seeing isn't proving.

Anthropic opened the Cowork VM to security sensors. Now you can see what Claude did. But which agent acted? Was it allowed to? Can you show an auditor? That's Atbash: DNA for AI agents.

01A verifiable identity
02A mandate it can't exceed
03A tamper-proof record of every action
Our Claude plugin is live today.LIVE
CISO rolling out Claude? DMs are open.Read on LinkedIn
PROGRAM MEMBERSHIPMEMBER
NVIDIA logo
NVIDIA
Inception
MEMBERAtbash AI
WORKING ONOpenShell
STATUSIn progress

Runtime + mandate.

AI agents are moving into banks, payments and critical systems. OpenShell governs how an agent runs. Atbash adds who the agent is and what it's mandated to do.

OPENSHELL · HOW IT RUNS
sandboxpermissionsnetworktools
ATBASH · WHO AND WHAT, AT RUNTIME
identitymandatered linesaudit trail no one can delete
Building it inside NVIDIA Inception.Read on LinkedIn
SUPPORTED RUNTIMEINTEGRATION
OpenAI logo
OpenAI
GPT agents
WORKS WITHOpenAI
GATED ATTool call
STATUSSupported

The model reasons. Atbash decides.

GPT agents call tools that send, pay and deploy. Atbash sits at the tool call and returns ALLOW, HOLD or BLOCK before it runs, judged against the mandate you set, whichever model chose the action.

01A verdict before every sensitive tool call
02A named approver on every HOLD
03A signed record the agent can't rewrite
Atbash judges the action, not the model behind it.See integrations
Capabilities

Every consequential action, decided and recorded

Most of what your agent does just runs. Only the irreversible actions get a verdict, and every verdict leaves proof.

◎

The Scanner

A read-only scan shows the authority your agent has that no workflow needs.

Learn more
support-agentExposure 92
CAN DO FOR YOU
✓ answer tickets
✓ refund small orders
NO WORKFLOW NEEDS
× email any domain
× export the CRM
⌁

Verdict feed

ALLOW, HOLD or BLOCK before each sensitive tool call. Fail-closed if no verdict arrives.

Learn more
ALLOWpost_summary0.4s
HOLDsend_bank_transfer1.2s
BLOCKexport_crm0.9s
ALLOWupdate_ticket0.3s
◐

Human approval

A HOLD pauses the agent until a named person answers. The answer goes into the record.

Learn more
HOLDfinance-agent
Transfer $25 to a new external vendor
red line: unverified payee
ApproveDecline
FLfinance-lead · notified
▤

Evidence

Signed, tamper-evident and kept off the agent's machine. Ready when the review begins.

Learn more
Agent CertificateISSUED
mandateFinance operations
ownerfinance-lead
red lines5 active
atb-cert:0x7f3a…c91e
How it works

From one hook to signed proof, in three steps

Connect the agent you already run, give it a mandate, and let every sensitive action pass through the boundary.

1

Connect your agent

Install the plugin for your framework, or wrap tool calls with the SDK.

2

Give it a mandate

State the job. Atbash proposes red lines from the gap, and you decide.

MANDATE
Keep the books reconciled and suppliers paid within approved limits.
RED LINES
Unverified payeeHOLD
Alter closed ledgerBLOCK
Authority Exposure92 → 34
3

Decide and record

Each sensitive call gets a verdict before it runs, and a record you can hand over.

VERDICT SUMMARY · 24H
1,284
ALLOW
12
HOLD
3
BLOCK
illustrative figures
Latest recordACC-…-4418
Who it's for

Built for whoever answers for the action

One boundary, three kinds of accountability.

Enterprise

Lead

For Risk and Compliance teams running agents in sensitive workflows.

Before execution.
✓Verdicts on treasury, deploys, data and orders
✓HOLD with a named approver
✓Signed Verdict Records per decision
✓Plugins for seven frameworks, SDK and API
✓Your cloud, private subnet or on-premises
Request a demo

Individuals

For anyone who gave an agent their inbox, files, photos or money.

92 → 34
✓Free, read-only scan
✓See your agent's blast radius
✓Set your own red lines
✓Lower your Authority Exposure
✓Keep an Agent Certificate
Scan your agent

Government

For public-sector teams running agents near classified and citizen data.

Chain of command.
✓Mandate, owner and approver per agent
✓Red lines on disclosure and recipients
✓Identity mismatch blocked and contained
✓Tamper-evident records for oversight
✓Deployed inside your infrastructure
Request a briefing
Want to see what your stack can safely automate?
Bring one agent and one workflow you worry about. We'll map the boundary with you.
Talk to us
Atbash Papers

Foundations of autonomous software

Short essays on authority, agents and the decisions between them.

What could your agent do to you?

Rolling out selectively to teams with agents in sensitive workflows. We want the conversation more than the signup.