Why we invest in Deception Engineering

We spend a lot of time thinking about what attackers will do next. Not just which vulnerability they will exploit, but how they work, how they find their targets, what they look for once they get in and how quickly they can move through an environment.

And that is changing. AI (and Agentic Ai) is already becoming part of the attacker's toolbox. Microsoft Threat Intelligence has seen threat actors use AI to create phishing content, translate messages, analyse stolen data and generate or debug malware. Google Threat Intelligence is also seeing actors move towards agentic AI and more automated attack workflows. Anthropic even documented a 2025 espionage campaign where a Chinese state-sponsored actor used an AI agent to perform large parts of attacks against around 30 targets, with limited human involvement.

So for us, this isn't something to worry about in five years. It is happening now, the good news is: There is a solution!

Why deception?

Deception usually means misleading someone, and that can sound negative. In security, however, that is exactly the point: we mislead the attacker, not your users.

The idea behind deception engineering is actually quite simple. We put something in your environment that looks interesting to an attacker, but shouldn't be touched by a normal user. This could be an identity, a credential, a server, a file or information about another system. If somebody touches it, we know something interesting is happening. But there is another part which we find even more interesting: what if the attacker isn't a person looking at your environment, but an AI agent?

An AI agent still has to make choices. It needs to discover systems, interpret information, select targets and decide where to go next. And the information it finds doesn't necessarily have to be real. That gives defenders an interesting opportunity.

Attackers automate, so should we..

Recent incidents show how quickly the attacker side is moving. Google observed an AI-enabled operation in 2026 where a compromised cloud resource was turned into an agent-enabled mass credential harvesting campaign in under six hours. Google has also reported what it believes was the first threat actor use of an AI-developed zero-day exploit.

At the same time, we still see very human attackers adapting quickly. ShinyHunters, for example, exploited an Oracle PeopleSoft vulnerability as a zero-day earlier this year. More recently they returned with modified exploitation techniques designed to bypass defensive WAF rules.

Attackers learn and adapt, and increasingly they can use AI to help them do that faster. Adding another thousand alerts isn't going to solve that. We need to detect the things that really matter and be able to respond quickly.

That is where we see Deception Engineering becoming really interesting.

Make them show their hand

Deception gives us a signal with context. If an attacker, human, automated or AI-driven, starts interacting with something that has no legitimate reason to be touched, we want to know. We can then feed that information into our CARs model: Collate, Analyse and Respond. We collate what we see, analyse what the attacker is doing and use that information to decide how to respond. That response could be human, it could trigger a Purple Team investigation or, when we have enough confidence, parts of that response could be automated as well.

We don't know exactly what an AI-driven attack will look like in three years, but we do know attackers are already experimenting with it today. So are we.

Let the attacker look around. Just make sure you're watching.