Even as powerful AI systems like Mythos raise fresh cybersecurity concerns, new research shows AI cyber defence is still far from reliable. A study by Simbian found that today’s most advanced AI models can attack systems better than many humans, but they still struggle badly when asked to defend them.
The study, led by Ambuj Kumar, tested 11 leading AI models in realistic defence scenarios. These models were asked to detect hackers hidden inside huge volumes of security logs. The results were alarming. Not one model passed the test. Even the top performer, Claude Opus 4.6, detected only a small share of real threats and missed entire attack categories.
Also Read : Oil jumps as Trump reviews Iran options
Simbian’s new cyber defence benchmark shows advanced
Simbian’s Cyber Defense Benchmark is one of the first serious tests built around real-world cyber defence. Instead of giving AI simple security questions, researchers dropped raw machine logs into the system and asked the models to find malicious activity on their own. This better reflects how real security teams work every day.
The biggest weakness in AI cyber defence is context. AI models are strong at offensive tasks because they attack known targets. Defence is harder because cyber defenders must detect threats from weak and incomplete signals. That means AI must reason through noisy evidence, ask smarter questions, and identify hidden patterns. Most models failed at this.
The findings highlight a growing cyber risk. AI-powered attackers are improving fast, while AI cyber defence still misses more threats than it catches. As advanced models become more widely available, security teams face rising pressure to build stronger AI defence systems before attackers gain a bigger edge.
Also Read : CBSE Class 12 results expected by May 23 evaluation process on schedule