Making AI Headlines · Page 9
Open Models Approach the Cyber Exam's Restricted Top Student
· New York time
DeepSeek and GLM models came close to OpenAI's most advanced cyber model in an experiment asking AI systems to break into simulated networks, WIRED reports. OpenAI's model is not public, while the other models are available more broadly. The comparison matters because a capability need not be the absolute best in a test to be consequential. Models that can be obtained and run widely may reach users who cannot access a restricted system. The experiment measured performance in simulated networks, not the outcome of every real-world attack. For defenders, the result is another reason to pay attention to how quickly capabilities spread. For model makers, a lead on a test is less comforting when competitors are close behind. The class has a top student, but the rest of the room is catching up without waiting for an invitation to the honors program.
Read this story in the full newspaper →Satire. Not factual reporting. Read the original reporting from the linked publishers.