What the Largest Controlled Study of AI Attackers Reveals About Cyber Deception
Cyber deception was built to fool human attackers. For decades, honeypots, honeytokens, and decoys have relied on assumptions about how adversaries recognize threats, select targets, and respond to suspicious environments.
But autonomous, AI-driven attackers don’t follow the same rules.
So, does cyber deception still work against AI?
Horizon3 researchers tested 21 AI models across 10 providers, analyzing 10,962 attacker decisions and comparing their behavior against 47 human red-teamers.
The findings reveal a surprising advantage for defenders:
- 2.1x more susceptible: AI attackers fell for deception 78.5% of the time, compared to 37% for humans.
- 73% recognition–action gap: Even when AI models explicitly recognized a trap in their reasoning, they attacked it anyway approximately 73% of the time.
- 21 out of 21 models: Every AI model tested fell for traps more frequently than the human baseline, including the most advanced frontier models.
Yet AI attackers were also better at identifying real vulnerabilities, correctly recognizing risks 77.6% of the time compared to 44% for humans.
The takeaway? AI attackers are more capable at finding real vulnerabilities, but they’re also more likely to trigger deception.
For security teams, this changes the role of cyber deception from diverting attackers to detecting them early.
Inside the Whitepaper
Discover what Horizon3’s research reveals about AI-driven attacks and how security teams can adapt their defenses.
Learn:
- Why AI attackers fall for deception more often: Explore how 21 AI models performed against human red-teamers across different types of deceptive artifacts.
- The recognition–action gap: Understand why even advanced AI models can identify honeypots and honeytokens yet attack them anyway.
- Which traditional deception strategies no longer hold: See why hiding decoys and diverting attackers may be less effective against autonomous agents.
- Why AI attackers remain a serious threat: Examine how models can identify legitimate vulnerabilities more accurately while remaining susceptible to deception.
- How to turn deception into early detection: Learn why honeytokens, canaries, and other tripwires can provide valuable warning signals against AI-enabled attacks.
- How to prepare for different AI attacker models: Understand why frontier models and self-hosted, open-weight agents require defensive strategies informed by their actual behavior.
The research also compares deception effectiveness across file systems, .htaccess files, HTTP responses, and HTTP requests, finding that AI attackers were more likely than humans to engage with planted bait in every tested category.
Who Should Read This
This whitepaper is designed for:
- Chief Information Officers (CIOs) and Chief Information Security Officers (CISOs)
- Security Architects
- Threat Detection and Response Leaders
- Security Operations and Engineering Teams
- Threat Intelligence and Active Defense Teams
- Security leaders preparing for AI-enabled and autonomous threats
Whether you’re already using cyber deception or evaluating how your defensive strategy should evolve, this research provides actionable insights grounded in observed attacker behavior.
Download the Whitepaper
AI attackers may be harder to stop, but they can be easier to detect.
Traditional deception strategies were designed to distract human attackers from valuable assets. Autonomous AI agents can engage with decoys while continuing to pursue real vulnerabilities.
That means security teams need to rethink how they deploy deception, using honeytokens, canaries, and tripwires as early-warning signals rather than relying on misdirection alone.
Download Hacking the Hackers: Can You Still Deceive an AI Attacker? to explore the research and discover how to adapt your deception strategy for the next generation of AI-driven threats.
