Can AI Actually Hack? Inside AI That Finds & Exploits Vulnerabilities

Описание: ▶ Full write-up, references & tools → https://secureinsecurity.com

For most of computing history, finding a serious, previously unknown security flaw (a zero-day) depended on scarce human talent: a researcher who could read unfamiliar code, form a hunch, test it, and keep going when the first ninety ideas failed. Turning that flaw into a working exploit took even rarer skill. In 2026, that scarcity changed. Frontier AI models can now read large codebases, point to where the bugs likely are, confirm them by running the software, and in a growing share of cases produce a working exploit, with little human guidance at each step.

Two models define the moment. Claude Mythos Preview, announced by Anthropic on April 7, 2026, reported finding thousands of high-severity zero-days across every major operating system and web browser, including a flaw that had sat in OpenBSD for 27 years. Anthropic decided not to release it publicly and instead launched Project Glasswing, a defensive coalition of major technology and infrastructure companies. OpenAI Astra, detailed on September 2, 2026, became the first OpenAI model rated "Critical" for cybersecurity under the company's Preparedness Framework, after it discovered and chained zero-days on its own.

So, can AI hack? The honest answer is yes. The more useful questions for a practitioner are: how good is it really, who has access, how fast can defenders absorb what it finds, and what goes wrong along the way? This document answers those questions for IT and security practitioners and for learners who want an accurate picture without the hype. It covers the timeline that led here, the two headline models, the capabilities AI systems now demonstrate, the shift in means and opportunity for both attackers and defenders, documented real-world misuse, the negative impacts worth taking seriously, and a practical checklist for what to do now.