Beta · 4 spots remaining

Secure AI agents before they act.

Runtime security that sits inside the agent loop and stops attacks before your agent acts on them.

Join the beta See the benchmark → Try to fool our agents
LATEST CRITICAL EVENT · 1H AGO SAMPLE DATAHIGH

Tool/Data Exfiltration on order_agent

LAYER · DATA  ·  POLICY REV 2

47%
CP-1 User message
9%
CP-2 Data retrieved
28%
CP-3 Tools called
16%
CP-4 Agent response
30 ATTACKS BLOCKED 2.2% block rate over 7 days.

Hidden instructions in a web page, email or document are caught while the agent reads them, before it acts on them.

0 of 642 attacks got through
52.0% secure task completion
0 of 62 benign episodes with an intervention
Design partner beta Open to teams running AI agents. 10 spots, 4 remaining.

Why now

Agents act now. Their guardrails weren't built for that.

Poisoned content used to mean a wrong answer. With tools, it means a wrong action.

01

Tool output is the attack surface

An agent that reads a web page, an email or an API response will follow instructions hidden in it. That is why prompt injection tops the OWASP list.

02

Today's guardrails force a trade

In our benchmark the open-source options either let attacks through or broke normal work. LlamaFirewall let 29.0% of attacks succeed and intervened on 38.7% of benign episodes. NeMo Guardrails intervened on 54.8%.

03

Blocking has a cost

A defense that breaks normal work gets switched off. Our benchmark measures both: attacks that got through and normal tasks blocked.

Where we sit in the market →

Design partner beta

Secure AI agents before they act.

Limited to 10 teams for 4 to 6 weeks. 4 spots remaining.

Join the betaFor investors →