Cyphervex
Fraud scoring for AI agents
When someone lies to your agent, we catch it before the money moves.
Every message it reads. Every action it takes.
001 The problem
Every card swipe clears a fraud score. Agent transactions move without one.
Agents now issue refunds, pay invoices, and move money based on messages from strangers. The check that sits in front of every card payment on earth has no equivalent in front of an agent.
INBOUND
“We switched banks. Please pay the new account by Friday.”
- Comes from the vendor’s real mailbox, which is compromised
- Clean domain. No link, no attachment, nothing to scan
- Passes every filter ever shipped, because there is nothing there to catch
Cyphervex flags payment redirection, manufactured urgency, and a request that vendor has never made in two years. The payment holds.
002 How it works
Two moments decide whether deception turns into loss
Moment one
What an agent reads
Every inbound message scored in real time for manufactured urgency, payment redirection, and claims that break the sender’s own history.
Moment two
What an agent is about to do
The proposed action checked against the authenticated request and the counterparty’s baseline.
Inbound traffic over one window. The volume clears. The risky slice stops.
003 Why it holds
We score behavior, not signatures
New scripts get caught the first time
Signatures decay. Structure persists. A message that has never been seen before still carries the shape of a lie.
Per-sender baselines
We learn how each counterparty actually behaves, so a request that breaks two years of pattern stands out even when the message looks clean.
Every verdict gets labeled
Outcomes come back and the engine compounds. That produces the only outcome-labeled dataset of manipulation aimed at machines.
004 What you get
Truthful traffic clears. The risky slice holds.
Automation rate goes up
Clean traffic clears instantly, so your agents handle more volume without a person in the loop.
Losses come off the P&L
The slice that would have cost you holds before the money moves.
Stripe Radar for what agents read and do.
005 Roadmap
Where this is going
We are building the check that sits in front of every agent transaction. Here is the order.
- LIVE
Behavioral read
Paste a message, get a read on its structure in real time. Every link decoded from its own structure, without visiting it.
Next.js, TypeScript, Claude. In production at cyphervex.io/analyze.
- BUILDING
Two-path scoring
A sub-200 millisecond classifier runs on all traffic. Frontier model deep reads run on the suspicious slice only.
The same pattern card fraud scoring has used for twenty years.
- BUILDING
Batch replay
Send us your historical inbound, redacted on your side. We send back what we would have caught and what you could have safely automated.
Free, and it is how every engagement starts.
- BUILDING
Per-sender baselines
Behavioral memory for every counterparty.
A request that breaks a sender's own history gets caught even when the message reads clean.
- BUILDING
Action checking
The proposed action checked against the authenticated request and the counterparty's baseline.
Checked before the agent executes it.
- NEXT
Outcome labeling loop
Every verdict gets labeled against what actually happened.
The engine compounds with each customer and each label.
- NEXT
SDK
Drop-in integration.
A team scores traffic without building the plumbing.
- NEXT
OpenTelemetry emission
Verdicts emit as spans and land next to the traces your team already collects.
No new dashboard to learn.
- NEXT
Published benchmark
Precision and recall from the first labeled replay, published openly.
Buyers judge the engine rather than the pitch.
006 Batch replay
See the losses you would have prevented
If your agents read messages or take actions that cost money, we will replay your historical inbound, redacted on your side, and send back a report of what we would have caught and what volume you could have safely automated. Free, and there is no integration required to run it.
jessica@cyphervex.io