AI Can Be Fooled. That's Why Security Still Needs Experts.

Summary

 

In a Dark Reading column from Sept. 11, 2026, security experts argue that AI governance can't wait. The lead example is GuardBreaker, a technique researchers attributed to the Russia-aligned group UAC-0099. The attackers embedded a nuclear-weapons prompt as a comment inside their malware. When an AI model was asked to analyze the file, its safety guardrails triggered and it refused to continue. The malicious code went unexamined.

 

As the expert puts it, rather than making malware more sophisticated, threat actors "manipulate AI's defensive reasoning to quietly compromise a victim's networks."

 

He also warns that AI is shrinking the time between finding a vulnerability and exploiting it from years to hours, which makes traditional patch cycles hard to sustain. His recommendation: put governance frameworks in place now, and back them with layered controls that combine detection, behavioral analysis, sandboxing, heuristics and human-led engineering, rather than relying on AI alone.

 

 

 

GuardBreaker is a warning for any organization whose security depends on AI alone. If one line of text can make a model stop looking, the model can't be your last line of defense.

 

That's why InfoSight's operating model is Powered by AI, Fueled by Experts:

 

AI surfaces the signal, and analysts make the call. In our 24/7 SOC, AI speeds up triage, and experienced analysts review what it flags and what it declines to judge. A refusal from a model is a reason to escalate, not a verdict.

 

Mitigator® gives you continuous vulnerability and risk intelligence. When the window between discovery and exploitation is measured in hours, you need to know your exposure now, not at the next quarterly scan.

 

Purple Team SOCaaS tests your defenses the way attackers operate, including techniques built to evade automated tools, so you find blind spots before an adversary does.

 

For organizations in financial services, healthcare and manufacturing, AI governance is quickly becoming a board and examiner question. Layered, expert-led security is how you answer it with evidence.

 

CTA: Want to know how your defenses would hold up against AI-evasion techniques? Talk to an InfoSight expert →