Artificial intelligence is making hackers faster, but that speed may also become one of their biggest weaknesses. As AI agents automate vulnerability discovery, reconnaissance and exploitation, cybersecurity defenders are finding new reasons to embrace an old strategy: deception.
Honeypots, fake credentials and decoy systems have been part of the defensive security arsenal for decades. Now, as autonomous AI agents race through networks and process information faster than human attackers ever could, deception technology may offer security teams a way to turn that automation against them.
That's one of the more intriguing possibilities explored in the latest episode of CYBR.HAK.CAST, where hosts Michael Farnum and Phillip Wylie spoke with Jacob Krell of Suzu Labs about AI-powered hacking, Capture the Flag (CTF) competitions and the changing relationship between attackers and defenders.
The conversation began with Krell's research into how AI is disrupting competitive hacking. But it quickly evolved into something with broader implications for enterprise security:
What happens when defenders deliberately feed AI-powered attackers information designed to mislead them?
Full episode and related article:
, and CYBR.SEC.Media
Bill Brenner
When AI hackers chase the wrong flag
The discussion turned to an experiment conducted by Justin Hutchins, who develops CTF challenges for CYBR.HAK.CON and other CYBR.SEC.Community events.
After learning about Krell's presentation, Death of the CTF, Hutchins reportedly stayed up late redesigning a competition to include false flags intended to trick AI agents attempting to solve the challenges.
The idea was straightforward. If AI systems can identify and capture legitimate flags with extraordinary speed, perhaps they can also be manipulated into pursuing convincing but worthless targets.
That experiment illustrates an emerging opportunity for cybersecurity defenders. AI agents can analyze enormous volumes of information, identify suspicious configurations and execute attack sequences with minimal human intervention. But their ability to process information quickly doesn't necessarily mean they can reliably distinguish legitimate targets from carefully constructed deception.
Experienced penetration testers may recognize contextual clues suggesting that something isn't what it appears to be. Automated agents, particularly those operating at high speed, may be more inclined to interact with whatever appears to satisfy their objectives.
That creates an opening for defenders willing to exploit the limitations of AI-driven reconnaissance and exploitation.
From static honeypots to active AI-powered defense
Krell believes deception technology remains underutilized in enterprise environments, particularly as organizations invest heavily in AI-powered security operations.
"In the enterprise, I definitely think honeypots are underutilized," Krell said.
His vision goes beyond traditional honeypots sitting quietly on a network waiting for someone to interact with them.
Imagine an endpoint detection and response (EDR) system identifying suspicious activity involving network shares. Instead of simply generating an alert, the system could automatically deploy a convincing decoy share containing information designed to attract the intruder.
That deceptive environment might include false credentials, misleading files or even prompt-injection content intended to influence an AI agent processing the material.
The objective wouldn't necessarily be to stop the attack outright. It could be to identify malicious activity earlier, divert automated reconnaissance or generate higher-confidence detection signals.
Krell sees such capabilities as part of a larger shift toward security systems that respond dynamically to suspicious behavior rather than simply waiting for analysts to investigate alerts.
"I definitely think it's going to go towards more active detection away from kind of the passive detection that we're currently running," he said.
These scenarios remain forward-looking ideas rather than documented enterprise deployments discussed in the episode. But they illustrate how AI and deception could potentially work together to reshape detection engineering.
Why deception could become more effective against AI agents
Wylie pointed to observations from experienced red-team operators who have encountered deception technology during penetration testing engagements.
Even skilled human attackers can be caught by carefully constructed decoys. When attackers move quickly through unfamiliar environments, seemingly legitimate assets can become attractive targets.
AI introduces another dimension.
An autonomous agent conducting reconnaissance may encounter thousands of files, configurations and network resources. If a defender can place convincing deceptive material in that workflow, the agent may interact with it before recognizing that something is wrong.
Farnum suggested that cloud infrastructure and automation could make those decoys increasingly dynamic, allowing defenders to generate realistic-looking systems as suspicious activity unfolds.
The possibilities range from simple decoy documents to more elaborate simulated infrastructure designed to resemble production environments.
There are limits, however. A fake target that works against one model or attack workflow may fail against another. Prompt injection can influence some AI systems, but its effectiveness depends on how agents process untrusted information and what permissions they possess.
Deception is therefore best understood as another layer of detection and disruption, not a guaranteed way to defeat autonomous attackers.
The advantage defenders keep forgetting
Near the end of the conversation, Krell identified a problem that could undermine even the most sophisticated deception strategy: Many organizations don't fully understand their own environments.
Defenders should possess one fundamental advantage over attackers. They know which systems, applications and resources are legitimate, while intruders must discover that information.
But incomplete asset inventories, forgotten infrastructure and decades of accumulated technology can erode that advantage.
As Krell observed, organizations frequently lack a complete picture of what they own and operate.
That creates problems for deception engineering. To build convincing decoys, security teams must understand the real systems those decoys are intended to imitate. They also need to distinguish legitimate activity from suspicious interactions with deceptive assets.
Otherwise, organizations risk introducing additional complexity into environments they already struggle to manage.
Making AI attackers work against themselves
The larger lesson from the CYBR.HAK.CAST discussion is that cybersecurity defenders don't necessarily have to match AI-powered attackers step for step.
Instead, they can exploit the attacker's uncertainty about the environment.
AI agents may conduct reconnaissance faster, identify potential vulnerabilities more efficiently and automate tasks that previously required hours of human effort. But speed can also magnify mistakes when those systems encounter misleading information.
For security teams, the opportunity lies in combining accurate asset inventories, strong detection engineering and deception mechanisms that make suspicious activity easier to identify.
The next generation of honeypots may not simply wait for attackers to stumble into them. They could appear dynamically, adapt to attacker behavior and exploit the assumptions embedded in automated attack workflows.
AI is making it easier for attackers to find what they're looking for. Deception technology could make it much harder for them to know whether they've found the right thing.
