Get in touch

Agent security

Internal guardrails for secure AI agents.

Every action is checked inside the model before it runs.

Works on open-weight models.

Built by researchers from Fraunhofer HHI.

Demo

Example: a coding agent reads a poisoned README.

Prompt injection

Do you notice anything problematic in this terminal output?

Agents read thousands of lines in seconds, faster than anyone can review. Layerwise reads the model itself and shows, automatically, what it intended at every step.

Offer

Test, trace and guard your agent.

The benchmark

We attack your agent with its own tools and show where your model fails.

  • Attacks matched to your agent and its tools
  • Measured against the checks you run today

The trace

Each action is traced to its sources and kept as evidence for an audit.

  • The intent and source of every flagged step
  • Kept for your security review

The guardrail

Reads the model's internal signals and blocks actions before they run.

  • Checked at every step of the agent
  • Holds when attackers adapt

Proof

We catch what guardrails miss, and keep catching it when attackers adapt.

Measured against the best available guardrail for each claim, on open-weight models.

More effective

0.98vs 0.87

Catches what external guardrails miss

Standard prompt injections, AUROC

Detection quality · 0.5 is chance

Layerwise0.98
DeBERTa-v3 guardrail0.87

Faster

10%vs 1,000%

Security without slowing the model

Extra compute per answer, against today's attribution research

Compute overhead · lower is better

Layerwise10%
Attribution research1,000%

Harder to attack

0.99vs 0.51

Stays secure when attackers adapt

Attacker trained against each detector, AUROC

Detection quality · 0.5 is chance

Layerwise0.99
Meta Prompt Guard 20.51

The number that matters

Any detector scores well on the attacks it has already seen. Ours holds when the attacker is trained against it: 0.99 where the best deployed guardrail falls to 0.51.

Benchmarks on AgentDojo.

Deployment

Deployment on your infrastructure.

Your infrastructure

It runs inside the serving of your model, on your servers or in your cloud.

Your data stays with you

Prompts, tool outputs and actions do not leave your environment.

Same endpoint for your agent

Your agent calls the same API as before. Nothing changes in its code.

Pilot

Your agent, secured in eight weeks.

Together with our scientists, we implement our technology and set up the security for your agent.

  1. Week 1

    Connect

    Use our endpoints, on-premises or in your cloud.

  2. Weeks 2 to 6

    Find and guard

    We show you your model's failures, then build the guardrails for your use case.

  3. Weeks 7 to 8

    Deploy

    Once we beat your current checks. You decide on rollout.

  4. After week 8

    Ongoing support

    In production, and when your agent, its tools or the attacks change.

What we need from you

  • The open-weight model your agent runs on

    Served on your machines or in your cloud.

  • One agent workflow and its tools

    Shown together with what it must not do.

  • One engineer as a contact

    For access and the endpoint switch.

What you get

  • Your model's failures

    Shown on your own agent and its tools.

  • Guardrails for your use case

    Measured against the checks you run today.

  • A record of every flagged step

    As evidence for your security review.

Let's secure your agent now.

Tell us which model your agent runs on and which tools it can use.

Get in touch