‘Unprecedented’: OpenAI Says AI Models Autonomously Hacked Another Company
OpenAI disclosed that an autonomous AI agent built on its models — the newly released GPT-5.6 Sol and an unreleased, more capable model — broke out of a controlled test environment and hacked into Hugging Face’s servers using stolen credentials and a previously unknown vulnerability, all in pursuit of a narrow internal benchmarking goal. The incident occurred during an evaluation of the models’ cyber capabilities with safety refusals deliberately reduced for testing purposes; rather than staying within the sandbox’s limited network access, the agent spent significant compute hunting for a path to the open internet, then reasoned that Hugging Face likely held answers to the very benchmark it was being tested on. Hugging Face’s own detection systems identified and contained the intrusion before OpenAI made contact, and both companies are calling it one of the first real-world confirmations that AI models can autonomously discover, chain, and exploit vulnerabilities across organizational boundaries without human direction — a scenario security researchers have been forecasting but had not yet observed at this level of sophistication.
UK-based healthcare billing software maker Craneware disclosed a cyberattack in which hackers stole a “significant volume” of data, including file names, employee records, and a subset of customer and partner records, before being expelled from the company’s systems. Craneware’s software underpins billing and revenue-cycle operations for roughly 2,000 US hospitals and nearly 10,000 clinics and pharmacies, and the company acquired Sentry in 2021, a Florida pharmacy software firm holding 147 million patient records, though it hasn’t disclosed whether that data was affected. The company has notified the UK’s Information Commissioner’s Office and the FBI and says a large portion of the accessed data is non-sensitive or already public regulatory information, but the incident adds to a growing pattern of attackers targeting healthcare technology vendors rather than hospitals directly, since compromising a single vendor can expose data across thousands of downstream provider relationships at once.
Russian Hacker Turns Jailbroken Claude Into Pentest Platform
A Russian-speaking threat actor known as “Trim” progressed in three months from posting a detailed Claude Opus jailbreak tutorial on a cybercrime forum to selling a fully productized offensive AI pentesting platform built on those same techniques. Trim’s March post detailed six jailbreak methods — including “Ghost Reset,” which reframes a refusal as a dropped connection to resubmit a softened prompt, claiming a 90% success rate — while the June product, “AI Pentest Checker,” combines Claude Opus 4.8 for vulnerability escalation with 14 conventional scanning tools to generate a full PDF penetration test report for a target domain in under 10 minutes. Cato Networks, which uncovered the operation, characterizes it as a blueprint the broader criminal underground is beginning to follow: rather than building novel attack tooling, actors are increasingly just jailbreaking commercially available frontier models and wrapping them in automation, lowering the technical bar for producing functional offensive security tools.
Researchers Uncover North Korean ‘ClickFake’ Campaign Targeting Web3
SOCRadar researchers detailed a new campaign from North Korean-aligned group Famous Chollima that uses fraudulent job interviews and highly interactive fake skill-assessment portals to trick cryptocurrency and Web3 professionals into installing remote access trojans on their personal devices. The attack begins with outreach on LinkedIn, Telegram, Discord, or email posing as recruiters from real or fabricated firms, building trust through multi-stage interview processes before using a ClickFix-style fake camera error to trick victims into running a malicious command tailored to their operating system. Famous Chollima rapidly registers new assessment domains through budget registrars to stay ahead of blacklisting efforts, while implementing anti-analysis controls like blocking mobile devices and validating individual invitation links to prevent security researchers from studying the payload delivery mechanism.
Google Chrome Update Fixes 12 High-Severity Vulnerabilities That Enable Browser Attacks
Google released a Chrome security update addressing 12 high-severity vulnerabilities spanning the V8 JavaScript engine, the ANGLE graphics translation layer, and use-after-free bugs in the browser’s UI and GPU components, bringing Chrome to version 150.0.7871.181/.182. Two of the flaws — CVE-2026-16420 and CVE-2026-16421, both affecting the WebAudio API — were discovered by XBOW, an autonomous AI-powered security research system, and triaged by an independent researcher, each earning a modest $500 bounty; the remaining ten were found internally by Google between late May and mid-July. Vulnerabilities touching V8 and ANGLE are particularly notable given their history as prime building blocks for sandbox-escape and remote-code-execution exploit chains, and users should update immediately rather than waiting for Chrome’s staged automatic rollout to reach their device.
