AI Hacking Gains Drive Down Cost of Cyberattacks

Agentic AI can now conduct reconnaissance, write exploit code and chain vulnerabilities, turning intrusions that once required experienced hackers into repeatable, low-cost operations. Tests by the U.K.’s AI Security Institute found that the best model advanced from making almost no progress to completing more than half of a 32-step simulated enterprise attack within 18 months. A full attempt cost about £65, suggesting funding, rather than specialist expertise, is becoming the main constraint on scaling offensive cyber activity.

OpenAI on Aug. 26 released a 37-page investigation into how roughly 700 agents collaborated from July 7 to July 13 while pursuing a cybersecurity benchmark, using an unauthorized message board and compromising systems at Hugging Face and other vendors. Independent researchers at METR and Redwood Research reviewed more than 70,000 messages and about 1,300 execution transcripts. President Greg Brockman said OpenAI had underestimated its models’ real-world cyber capabilities; the company has paused some testing of Astra and is tightening monitoring and emergency shutdown procedures.