Technology / 08 Nov 2026
Free AI models are scary good at hacking
Vals, an AI measurement company, found that free open-weight models from DeepSeek and GLM are almost as capable as the best closed models at carrying out multi-step hacking tasks on simulated corporate networks.
Sources.
Free Chinese AI Models Are Getting Scary Good at Hacking | WIREDMenuWIREDMenuWIREDAccountAccountChevronExpandSearchCommentLoaderSave StoryCommentLoaderSave StoryTriangleXXLargeChevronFacebookXPinterestYouTubeInstagramTiktokwww.wired.comMedia credits.
Wikideas1 (CC0)commons.wikimedia.orgIvan Chumak (Pexels License)www.pexels.comflutie8211 (Pixabay Content License)pixabay.comModem kg (Pexels License)www.pexels.comContext and limitations.
The percentages and costs are Vals's own test findings on simulated networks, not independent real-world results. OpenAI did not respond to comment. Experts still debate how much AI changes cybersecurity fundamentally.