TLDR: New research from Forescout Vedere Labs indicates that while AI-driven cyberattacks are a future concern, current generative AI models exhibit significant technical shortcomings, failing to reliably execute end-to-end cyberattack pipelines. The study, involving over 50 simulations, found high failure rates in vulnerability research and exploit development tasks, suggesting AI remains a supplementary tool for threat actors rather than a primary enabler of major breaches.
A recent study by Forescout Vedere Labs has shed light on the current capabilities and limitations of generative artificial intelligence (GenAI) in the realm of cyberattacks, concluding that AI-driven hacking remains a future concern rather than an immediate danger. The research, published on July 10, 2025, involved over 50 simulations designed to evaluate how advanced large language models (LLMs) perform in two critical cybersecurity tasks: vulnerability research (VR) and exploit development (ED).
The findings highlight significant technical shortcomings across a range of models, including commercially available, open-source, and those found in underground communities. In the vulnerability research simulations, a substantial 48% of the assessed models failed the first task, with 55% unable to complete the second. The failure rates were even more pronounced for exploit development, where 66% failed the first ED task, and a striking 93% failed the second.
Beyond the high failure rates, a critical observation was the inherent instability and inconsistency of these AI models. During testing, LLMs frequently produced varying results even when presented with identical prompts, often timing out or generating content that was not usable for security purposes. Crucially, the research indicated that no artificial intelligence model tested was able to complete all required phases of a typical exploit pipeline autonomously. This means that, at present, the technology lacks the reliability and coordination necessary to carry out a complete, end-to-end cyberattack independently.
While commercially available LLMs generally performed better than their open-source and underground counterparts, even among these, only three models were able to produce a working exploit in the controlled scenarios. The study also addressed the concept of ‘vibe hacking,’ concluding that the risk posed by such attacks remains largely theoretical.
Also Read:
- Cyber Threat Intelligence Under Siege: How AI-Generated Fake Data Undermines Security Systems
- Accenture and Microsoft Deepen Partnership with Gen-AI Cybersecurity Initiative
In summary, the current analysis concludes that despite media attention and speculation, existing AI models are predominantly supplementary tools for cybercriminals rather than primary enablers of major breaches. This suggests that threat actors continue to rely on established methods and techniques, with a significant shift to fully AI-driven attacks not yet materializing.
