Agents act faster than your controls.
They read documents, call tools and move data inside workflows, in your cloud and on every laptop. The same few attacks reach all three.
AI agents now read your documents, hold your credentials and act in your core systems. Zenstra decides every action before it runs, and keeps the record that proves it.
Agents work inside workflows, in your cloud and on every laptop. Take the core apart to see where each attack is stopped, and what lands on the record.
They read documents, call tools and move data inside workflows, in your cloud and on every laptop. The same few attacks reach all three.
Each action is checked against the agent’s identity, the task it was given and your policy, inline, before it runs.
deny tool.call when outside(task.scope)hold payments.* for human approvalallow crm.read for member in task
Inline · p95 18 ms · Fails closed
Blockhop 7 · crm.export → outside email · not in task
Blockmcp core-banking · accounts.transfer · not in grant
Prompts and payloads never leave your VPC
Blocklaptop-114 · cursor · read ~/.aws · session contained
Each entry is chained to the one before it, so any change shows. Export it as evidence for your auditors and examiners.
#48213 block POL-114 v7 prev 9f2c…a71e#48214 hold POL-021 v3 prev c07a…5d12#48215 allow POL-003 v12 prev 4be1…09c3
Hash-chained · Exported as PDF and CSV
See the decision recordAgents run in workflows, in your cloud and on laptops, connected to your systems by credentials most identity and data tools never see.
No-code flows, copilots and coding agents are built by business teams and developers, not provisioned by IT.
Tokens, API keys, OAuth apps, service accounts and MCP connections, many with standing, broad access.
They decide and act in seconds, inside the systems that move money and member data.
Five public incidents from 2025, each with its source, and the Zenstra control that applies to that kind of action.
Hidden instructions in an email could lead Copilot to disclose internal data with no action from the user. Microsoft fixed it (CVE-2025-32711) and reported no exploitation.
During a code freeze, Replit’s agent ran destructive commands against the production database of a SaaStr founder’s project, despite instructions not to change anything.
Version 1.84.0 of the VS Code extension contained a prompt telling the agent to wipe systems and cloud resources. A syntax error stopped it from running.
Tokens held by an AI chat integration were used to export Salesforce data from customer instances between 8 and 18 August, and to search it for further credentials.
An npm package posing as a Postmark MCP server added one line in version 1.0.16 that copied every email an agent sent to the attacker.
Zenstra ties every agent to the person it acts for, maps every credential it holds, and narrows each grant to the task in front of it.
Most tools see an agent after it acts. Zenstra is in the path of every action, with the context to tell a helpful agent from a hijacked one.
Zenstra sits in the path of the action. Allow, hold or block is decided before the tool runs, not flagged in a log afterwards.
Each action is checked against the task the agent was given, so a hijacked agent holding valid credentials is still stopped.
The gateway runs in your VPC. Prompts and payloads never leave your cloud, and nothing in your agents is rewritten.
Every decision is chained to the one before it, with the policy version that made it, and exports as a report.
Agents, MCP servers and credentials across your platforms, in one inventory.
Learn more DecideAllow, hold for a person, or block, against identity, task and policy.
Learn more DetectBaselines per agent, and one agent stopped on one device in seconds.
Learn more ProveA hash-chained record your auditors and examiners can verify.
Learn more