Your company is starting to run AI agents that read email, use chat tools, open tickets and take actions. We attack them the way a criminal would, and prove whether they can be talked into something they must never do.
An agent can be socially engineered the way a person can. The difference is that it never gets tired, never gets suspicious, and holds real permissions inside your systems.
Agents built on Muse, OpenClaw, Hermes or your own code. Agents that read email, use chat tools, open tickets, run tools and take actions inside your systems.
Live today: the web, its API and Telegram. Coming: email, Slack, Microsoft Teams, WhatsApp, Discord and voice. The attacker builds a lasting identity and works the agent over days, not in one prompt.
You write the rules. We try to make the agent pay, change an account, disclose data, book something or send a message it must never send. Every rule points at a real tool the agent holds.
Every tool call the agent makes is recorded and checked against your rules. You see the transcript, the action and the rule it broke, graded critical to low. Nothing real is ever touched.
We test what your agent can be made to do, whatever takes control of it. There are three ways in. The harm always goes through the same tools and permissions, so that is where we look.
Someone talks the agent into it. A fake boss, a made up emergency, a request that sounds routine.
A message, a document or a tool result carries orders. The agent follows them as if they came from you.
Poisoned training data teaches the model a secret trigger. When it sees the trigger, it obeys the attacker instead of you.
Five steps, from registering the agent to the retest. The attacker runs on our own hardened server, reaching only the model provider and the targets you authorise.
Which agent, which channels it listens on, which actions it can take: payments, account changes, disclosures, bookings, messages.
Plain rules per agent. MUST NOT refund above a limit without approval. MUST NOT change a payee. MAY look up an order. Signed rules of engagement before anything runs.
Pretexts, urgency, poisoned documents, cross channel pressure, over days. Three test modes: through the agent's live channels, inside your own staging sandbox, or inside the DeepBlocker Agent Arena, a controlled world of hostile customers and dishonest websites where nothing real can be touched.
The sandbox records every tool call. Our verifier checks each one against your rules and grades the finding. A kill switch stops the campaign at any moment.
Report, event log, readiness score, and a rerun after every model, prompt, tool or permission change.
The same deliverable our voice fraud clients already buy, in the format their auditors already accept.
Trust companies, fund administrators, private banks and wealth managers putting agents in front of client and payment systems.
Teams that already test phishing and voice, and now have a third population of targets that never gets tired and never gets suspicious.
Agents that reset passwords, change MFA, open tickets and answer customers are the new help desk, and the help desk is where real attackers go first.
You pay for how much testing we run, not for how many agents you have. All prices exclude VAT and GST.
We are taking a small number of design partners now, at reference pricing, in exchange for a case study.