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.
A verdict before every sensitive tool call, and a signed record after it. Built for the teams who answer for what their agents do.
Runs where you build
Plugins for seven agent frameworks gate each tool call at the hook you already have.




Red lines you own
Proposed from the gap between mandate and reach. The agent can't edit its own.
From mandate to record
Every governed action walks the same path, and leaves the same proof.
Runtimes decide how an agent runs. Atbash decides who it is, what it may do, and keeps the proof.
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.
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.
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.
Most of what your agent does just runs. Only the irreversible actions get a verdict, and every verdict leaves proof.
Verdict feed
ALLOW, HOLD or BLOCK before each sensitive tool call. Fail-closed if no verdict arrives.
Learn moreHuman approval
A HOLD pauses the agent until a named person answers. The answer goes into the record.
Learn moreEvidence
Signed, tamper-evident and kept off the agent's machine. Ready when the review begins.
Learn moreConnect the agent you already run, give it a mandate, and let every sensitive action pass through the boundary.
Connect your agent
Install the plugin for your framework, or wrap tool calls with the SDK.
Give it a mandate
State the job. Atbash proposes red lines from the gap, and you decide.
Decide and record
Each sensitive call gets a verdict before it runs, and a record you can hand over.
One boundary, three kinds of accountability.
Enterprise
LeadFor Risk and Compliance teams running agents in sensitive workflows.
Individuals
For anyone who gave an agent their inbox, files, photos or money.
Government
For public-sector teams running agents near classified and citizen data.
Short essays on authority, agents and the decisions between them.