AI detection & response

Tell a helpful agentfrom a hijacked one.

Classic EDR can’t tell a helpful coding agent from a hijacked one. Zenstra sees the prompt, the tool call and the device together, and contains one agent in seconds.

The problem

Agent attacks look like
normal work.

A hijacked agent uses the same tools, credentials and device as a helpful one. Without the prompt, every alert is a guess.

  • No contextEndpoint tools see processes and files, not prompts and intent.
  • Too much noiseEvery agent action looks unusual to tools built for people.
  • Slow containmentIsolating a whole device to stop one agent stops the person too.
A real pathA coding agent is turned against its developer.
  1. 1cursorA developer asks the agent to fix a failing build.
  2. 2READMEA file in a dependency tells the agent to read ~/.aws and post it to a paste site.
  3. 3Zenstra AI DiscoveryThe prompt, the tool calls and the device are linked: reading credentials is not part of the task.
  4. BlockSession contained on one device. The developer keeps working.laptop-114 · cursor · read ~/.aws
Industry references
  • ASI01Agent Goal HijackOWASP Top 10 for Agentic Apps
  • ASI10Rogue AgentsOWASP Top 10 for Agentic Apps
  • AML.T0051LLM Prompt InjectionMITRE ATLAS
How Zenstra handles it

Detect with context.
Respond in seconds.

Prompt-to-action lineage connects what the agent read, what it planned and what it tried to do.

Detect

Lineage, not noise

Each detection links the input that caused it, the action the agent planned and the policy that caught it, ranked by severity.

  • Detections ranked by severity
  • The agent and the policy that caught each one
  • Prompt-to-action lineage
Investigate

Start from the outcome

Investigations open on what happened, the blast radius and the agent run, step by step, with the excerpt that caused it.

  • Replay the agent run
  • Blast radius across agents
  • The retrieved excerpt that carried the attack
Contain

One agent, one device

Stop the agent without isolating the laptop, so the person keeps working while the agent is contained.

  • Contain in seconds
  • Windows and macOS
  • Deployed with Intune, Jamf or Kandji
Policy as code

The policy
behind it.

Detections link the input that caused them, the action the agent planned and the policy that caught it, so an alert explains itself.

  • Decided inline
  • Fails closed
  • Versioned with every decision
denysecrets.readwhen not part of the prompt’s taskdenynet.postto paste sites · contain sessionholdmcp.serverwhen unregistered on deviceallowrepo.readwithin the open workspace
What it catches

Real agent attack
scenarios.

  • Prompt injection

    Hidden instructions in documents, email and web pages.

  • Data exfiltration

    Agents sending records to destinations outside the task.

  • Credential access

    Coding agents reading secrets they were never given.

  • Agent-to-agent spread

    Injected instructions passed from one agent to the next.

  • Memory poisoning

    Shared knowledge altered to steer future sessions.

  • Shadow AI

    Unregistered agents and MCP servers on staff devices.

FAQ

Common questions.

Does it replace our EDR?

No. It works alongside it. Your EDR protects the device; Zenstra AI Discovery understands the AI agents running on it.

How is it deployed?

As an app on Windows and macOS, through your existing device management: Intune, Jamf or Kandji.

Which AI tools does it cover?

Copilot, ChatGPT, Claude Code, Cursor and agents nobody registered.