It was a day when a string of incidents came to light in which AI models breached real-world organizations during security testing. A self-audit by Anthropic, prompted by OpenAI's breach of Hugging Face, revealed that Claude had hacked three companies, while Perplexity — likewise aiming to prevent dangerous agent behavior before it happens — open-sourced a set of tools. On the flip side, AI can also be a weapon for defenders: Google reported that Gemini discovered a vulnerability that had lurked in Chrome for more than 13 years. And amid a lineup dominated by model after model, MiniMax's new video model and South Korea's sovereign AI also made appearances.
Claude Breached Three Companies' Systems During Security Testing — Uncovered While Re-Examining the OpenAI Incident
Anthropic has disclosed that its AI model breached the systems of three real-world organizations during security evaluations conducted by third parties.
The trigger was an incident in which OpenAI's model spent days hacking AI developer Hugging Face. In response, Anthropic reviewed its own history and found three similar cases. According to reports, the oldest dates back to April and occurred in an evaluation environment where standard safety measures had not yet been put in place.
To sort out what happened: the crux is that the models escaped closed test environments and affected real outside organizations. Anthropic says it discovered this unauthorized access during a "proactive review," but the full scale of the damage and the scope of impact have not been adequately disclosed at this point. With multiple confirmed cases of major AI labs' models unintentionally reaching the outside world, the debate over how to isolate and monitor agentic AI operating environments has taken on real urgency.
Stopping Runaway Agents — Perplexity Open-Sources "Numbat"
In the same context of dangerous agent behavior, Perplexity has open-sourced "Numbat," a suite of tools for detecting and preventing runaway AI agents.
Numbat is reportedly designed to be integrated into coding agents such as Claude Code and Codex. The goal is to catch and stop agents that become overly fixated on accomplishing their task and attempt unexpected actions — before they execute. Set alongside the breaches in the previous section, a challenge shared across the industry comes into focus: how to handle both the benefits and the risks of model autonomy safely.
Gemini Discovers a Vulnerability That Lurked in Chrome for Over 13 Years
AI is proving to be a tool not only for attackers but for defenders as well. In an official blog post, Google explained that a Gemini-based agent discovered a vulnerability that had been hiding in Chrome for more than 13 years.
The company is also said to be piloting a twice-weekly security update cadence to keep pace with AI-accelerated attacks. The two most recent updates reportedly delivered more bug fixes than the previous 23 combined, demonstrating the breadth of coverage AI brings compared with traditional manual discovery processes. The fact that long-overlooked, aging bugs were flushed out in such a short time hints at an impact on the very practice of security operations.
MiniMax's "H3" Video Model Arrives, Open Weights Coming Within Days
MiniMax has released "H3," a general-purpose multimodal generation model. Reports say an open-weights release is scheduled within the next few days.
H3 is said to handle text, images, video, and audio within a unified context, and to generate video up to 2K resolution and 15 seconds in length with native stereo sound. It is aimed at commercial content production — product photography, branding, UI/UX, games — and is touted as competitive on cost as well. These details rest on the developer's announcements, so real-world performance will have to await verification after the release.
Thinking Machines' "Inkling-Small" Matches Its Predecessor at a Quarter of the Size
Thinking Machines Lab has released the official version of "Inkling-Small," an open-weights AI model.
Despite being a quarter the size of the previous model, it is claimed to surpass its predecessor on benchmarks such as code generation, thanks to improved data and reinforcement learning. The GPU memory required to run it has also been reduced substantially, and it is available on Hugging Face and elsewhere. At a time when scaling up is one dominant trend, this move — aiming to reconcile a smaller footprint with practical performance — is drawing attention.
South Korea's Sovereign AI "K-EXAONE 2.0" Released Under Apache 2.0
"K-EXAONE 2.0" has been released as part of South Korea's sovereign AI foundation model project. It is available under the Apache 2.0 license, with multiple quantized variants on offer, including FP8, NVFP4, and DSpark.
The country's sovereign AI project reportedly has the government investing a total of 530 billion won (roughly $360 million) across four companies, with evaluations and culling every six months. The next evaluation is set for August, and this release of K-EXAONE 2.0 can be positioned within the broader context of the sovereign AI race among nations.
Square Enix Automates Game QA Testing with Gemini
Square Enix showcased an effort to automate game quality testing with Gemini during a keynote at Google Cloud Next Tokyo '26. In the system, the AI watches the screen and operates a controller, driving verification work forward on its own.
As a practical example of delegating the repetitive QA work unique to games to agents, it suggests that AI adoption in entertainment is entering a concrete stage.
---