More attempts
Agents map exposed systems, try leaked passwords and send phishing around the clock. How much an attacker can do no longer depends on how many people they have.
AI-vs-AI defense
Attackers have AI agents now. Your defense needs them too.
An attacker with AI agents can probe thousands of companies at once, write a convincing lure for each one, and switch tactics as soon as one stops working. A team that reads every alert by hand can't keep that pace. Defense now needs automation that responds while the attack is happening and keeps up as it changes. That's how we work: the investigation runs at machine speed, the safe steps happen on their own, and a person stays on every decision that changes your systems.
- Notify. Posted the case summary to the security channel in Slack.
- Track. Opened a Jira issue for the finance team with the evidence attached.
- Enrich. Looked up the attachment hash in VirusTotal and added the result.
- Contain. Isolate LT-0442. Current state shown, waiting for a second person.
Approve isolationDecline
An attack, start to finish
See how we keep up with an attack driven by AI.
From a password spray to a contained incident and the write-up that follows.
What the video shows
- An attack no longer waits for a person at a keyboard. AI agents try thousands of passwords, write a lure for every inbox and switch tactics the moment one gets blocked. Here a password spray and a flood of sign-in prompts are blocked, and a reply inside a real invoice thread gets through.
- The tools that help your team help attackers too: more attempts around the clock, faster changes when one approach is blocked, and lures written in a colleague's style about a project that's really happening.
- The slow part is gathering, not deciding. Pulling the sign-in history, the device activity, the mailbox rules and the earlier cases together by hand is where the hours go, and it's the part an attacker's agents outrun.
- Attackers have AI agents now. Your defense needs them too.
- Every alert, investigated as it arrives. Steward gathers the accounts, devices, mail and cloud activity, links earlier cases and tests the harmless explanation first, so an analyst starts from a case, not a raw alert.
- It follows the attack as it moves. A new inbox rule, an app granted mailbox access, a sign-in from a new address and a lure sent to suppliers all join the open case, and the investigation picks up from what's new instead of starting over.
- Agreed ahead of time, run when it counts. You choose a short list of actions we can take straight away, like ending every session on a taken-over account. Each one still needs a second person, while the safe steps, like posting the case and opening a ticket, are already done.
- Fast is no use if it's a guess. When a supplier's real mailbox asks to change bank details, the evidence can't settle it, so Steward stops and hands the case to our analysts on shift, with the open questions attached.
- Detections that keep up. What we learn from each case goes back into the detections, so the next attempt gets caught sooner, even with new addresses, different accounts and a rewritten lure.
- The agents on your side need limits too. An attacker hides an instruction in an email your assistant reads. Omni gives the agent only the tools that case needs, so the request to turn off sign-in checks is refused and recorded.
- Already an incident? Our incident responders lay out what the attacker's agents did, in order, with where each fact came from, and write it up so your leadership, your insurer and your auditors can follow it.
- AI for the speed. People for the decisions.
- See how fast a case moves on your own alerts. Bring a few recent alerts and we'll work them through Steward with you.
What changed
An attack no longer waits for a person at a keyboard.
The tools that help your team write code and answer email help attackers too. The difference shows up in three places.
Faster changes
When one approach gets blocked, an agent tries the next one. A rule written for last month's attack can be out of date before anyone reviews it.
Better lures
Messages written in a colleague's style, in any language, about a project that's really happening. The spelling mistakes people were taught to look for are gone.
Where defense falls behind
The slow part is gathering, not deciding.
Before anyone can say whether an alert matters, someone has to pull the sign-in history, the device activity, the mailbox rules and the earlier cases, then put them side by side. Done by hand, that's where the hours go, and it's the part an attacker's agents outrun.
Hiring more analysts helps for a while. It doesn't change the shape of the problem when the other side can add agents for the price of compute.
How we keep up
Respond as it happens. Adapt as it changes.
Our service and our software split the work the same way: everything that can safely run at machine speed does, and everything that changes your environment goes past a person.
- Every alert, investigated as it arrivesSteward starts on each alert straight away. It gathers the related accounts, devices, mail and cloud activity, links them to earlier cases and tests the harmless explanation first, so an analyst starts from a case, not a raw alert.
- It follows the attack as it movesAttacks driven by agents change course quickly. When new activity shows up on an account, device or address already tied to an open case, it joins that case, and the investigation picks up from what's new instead of starting over.
- Response in real timeThe steps you've approved in advance run the moment the evidence supports them, not when someone gets to the queue. That's what keeps a fast attack from turning into a long incident.
- Detections that keep upWhat we learn from each case goes back into the detections, so the next attempt that looks like it gets caught sooner, even when the attacker has changed the details.
- It stops when the evidence is thinSpeed is no use if the answer is a confident guess. When the evidence is weak or contradicts itself, Steward stops and hands the case to an analyst instead of closing it.
- Safe steps run on their ownAutomaton posts the case to the right channel, opens the ticket and checks the file or domain, without waiting for anyone to get to it.
- Containment agreed ahead of timeYou choose a short list of actions we can take straight away, like isolating a laptop at night. Each one still needs a second person to approve it, so speed never means one person acting alone.
- Analysts on shift around the clockOur team works the cases that need judgment, at any hour, and brings you in when there's a decision only you can make.
Your own AI
The agents on your side need limits too.
An AI agent that can query your EDR and your identity provider holds the same access an attacker is looking for. If it runs on a permanent key, a stolen prompt or a stolen laptop can turn your defense into someone else's tool.
Explore Omni- Access for one caseOmni gives an agent only the tools one investigation needs, with an expiry and a call limit.
- No keys to stealVendor and model keys stay inside Omni. The agent never sees them.
- A record of every callEach call is tied to a person or agent, a grant and a case, so you can see exactly what your AI did.
Questions
What people ask about AI-vs-AI defense.
Do you use AI to fight AI?
Yes, for the parts where speed matters and a mistake is cheap to check: gathering evidence, linking activity and testing explanations. A person reviews the result, and anything that changes your systems waits for approval.
Can it act on its own when an attack is moving fast?
Messages, tickets and lookups run on their own. For containment, you agree up front which actions we can take straight away. Those still need a second person to approve them, and nothing outside that list happens without you.
Do we have to replace our security tools?
No. We read from the endpoint, identity, email and cloud tools you already run, through your own Omni. Your logs stay where they are.
Can our own team use the software?
Yes. Steward and Omni are available on their own for teams that run their own SOC, and Automaton is in early access.