How it works

Four checkpoints inside the agent loop.

Prompt injection is the top risk on OWASP's Top 10 for LLM applications. Our checkpoints stop it where it enters, each one watching a single kind of content.

CP-1

User message

Checks what people send to the agent. Stops jailbreaks and disguised instructions.

Watches · What people send
CP-2

Data retrieved

Checks what the agent reads from tools, files and web pages. This is where hidden instructions enter.

Watches · What tools return
CP-3

Tools called

Checks each tool call the agent makes.

Watches · Tools called
CP-4

Agent response

Checks what the agent says back.

Watches · Agent response
In the benchmark

CP-2 did most of the work, and CP-3 caught 6 attacks the data checkpoint had passed. Everything is logged, so you can see every attack, set the rules for each check, and review every flagged event.

Why it compounds

Every verdict makes the next detection better.

SHIPPING

Wired into the agent loop, so it sees everything the agent does.

SHIPPING

Every verdict on a detection is training signal for the next one.

NEXT

Detectors trained on that signal. Not built yet.

PUBLISHED

Head-to-head results with competitor configs, so a buyer can reproduce them.

How we think about defensibility →

Next

See the attacks we run ourselves.

Our offensive engine finds the break before an adversary does.

Red team →Join the beta →