AI coding & desktop agents

AI on every laptop,contained in seconds.

Copilot, ChatGPT, Claude Code and Cursor work with source code, credentials and production access on staff devices. Zenstra AI Discovery sees the prompt, the tool call and the device together.

The risk

Coding agents hold
the keys to production.

A coding agent can read secrets, run shell commands and push changes. Hijacked, it does all that with the developer’s own access.

  • Secrets on diskCloud keys and tokens a coding agent can read.
  • Local MCP serversTools added to agents on a laptop without review.
  • Invisible to EDREndpoint tools see a process, not what the agent intended.
A real pathA coding agent reaches for the keys.
  1. 1claude-codeThe agent helps a developer debug a service on their laptop.
  2. 2~/.awsIt tries to read cloud credentials it was never asked to touch.
  3. 3Zenstra AI DiscoveryThe prompt, the tool call and the device are seen together, on the device.
  4. BlockSecret access blocked. Production shell needs a step-up.claude-code · read ~/.aws
Industry references
  • ASI05Unexpected Code ExecutionOWASP Top 10 for Agentic Apps
  • AML.T0112.000Local AI AgentMITRE ATLAS
  • AML.T0055Unsecured CredentialsMITRE ATLAS
Policy as code

How it is
stopped.

Rolled out with Intune, Jamf or Kandji. Stop one agent on one device without isolating the laptop.

  • Decided inline
  • Fails closed
  • Versioned with every decision
denysecrets.read~/.aws, ~/.ssh · unless in taskholdshell.execin prod · step-up requireddenymcp.serverwhen unregistered on deviceallowrepo.readwithin the open workspace
How Zenstra handles it

Device, prompt and action,
seen together.

Zenstra AI Discovery runs on Windows and macOS and is deployed through your device management.

See

Every AI on every device

Copilot, ChatGPT, Claude Code, Cursor and agents nobody registered, found on Windows and macOS.

  • AI tools and agents inventoried
  • Local MCP servers found
  • Deployed with Intune, Jamf or Kandji
Understand

Help or hijack?

Prompt-to-action lineage connects what the agent read to what it tried to do, so a helpful agent is not treated like a hijacked one.

  • Prompt-to-action lineage
  • Ranked detections
  • Context, not noise
Contain

One agent, one device

Stop the agent and keep the laptop working. The developer carries on while the agent is contained.

  • Containment in seconds
  • Step-up for production actions
  • Every action on the record
What it stops

Risks that start
on a laptop.

  • Hijacked coding agents

    Injected instructions in a repo or issue.

  • Secrets read by AI

    Agents reading ~/.aws, tokens and env files.

  • Secrets pasted into AI

    Staff sharing keys and data with chat tools.

  • Rogue MCP servers

    Unreviewed tools added to local agents.

  • Production shell access

    Shell and deploy actions held for step-up.

  • Shadow AI

    Unregistered agents on staff devices.

FAQ

Common questions.

Will developers notice it?

Routine work is allowed inline. Only actions outside policy are held or blocked, with a clear reason.

Which operating systems?

Windows and macOS, deployed with Intune, Jamf or Kandji.

Does it replace our EDR?

No. It complements it by understanding the AI agents on the device.