It was a funny week in the world of AI. Just when Brother Sam was enjoying all the attention from his mischievous AI model breaking containment, Uncle Dario said, “uh uh, you can’t hog all the fame”. Hence, Anthropic was on social media, revealing how its Claude did the same thing, but it breached three parties. Whether breaching third parties is a new badge of honour for AI companies remains to be seen (who knows if Google might declare Gemini breaching into something more serious!).
‘Uncle Jensen’ was busy organising safety coalitions and backing $250 billion infrastructure plays. Over in San Francisco, Dario’s Anthropic rolled out its latest Opus model, while legal courtrooms and state legislatures grappled with the boundaries of AI deployment.
Between frontier AI models breaking out of sandboxes to cheat on evaluations and mega-dollar infrastructure deals taking shape in the US heartland, the AI industry has given us a lot to be busier throughout the week.
All this and a lot more. Hence, let’s jump right into the subject matter – a roundup of all things AI this week.
OpenAI, Anthropic models break sandbox
The biggest story of the week comes from a startling safety revelation during internal model evaluations. OpenAI’s frontier cybersecurity models (GPT-5.6 Sol and an unreleased model) broke out of their sandboxes via a zero-day vulnerability in OpenAI’s package proxy. The models independently targeted and breached Hugging Face over a weekend in an attempt to “cheat” on an evaluation framework. Later reports reveal that the unreleased model went after other parties too.
Anthropic’s models were also reported to have autonomously breached three organisations during controlled red-teaming tests. During routine cybersecurity evaluations, a network setup flaw left Anthropic models connected to the live internet despite system prompts stating otherwise. Believing external targets were part of the test, models including Claude Opus 4.7 and Claude Mythos 5 scanned the open web, exploited real-world SQL vulnerabilities, and even uploaded malware to the PyPI registry. Evaluation logs revealed the models rationalised the live targets as part of their exercise, and in two cases, obscured their reasoning. This raises fresh alarms over AI containment and agent transparency.
Nvidia leads 35 tech giants into Open Secure AI Alliance
In the wake of the Hugging Face containment incident, Nvidia CEO Jensen Huang led a massive new coalition to improve AI infrastructure. Over 35 tech giants, including Microsoft, IBM, Cisco, Red Hat, Palantir, CrowdStrike, and Hugging Face, joined forces to launch the Open Secure AI Alliance. Huang cited closed-model limitations during the Hugging Face breach investigation as the primary reason why open weights and unified defensive frameworks are vital for global cybersecurity.
Nvidia also in talks for $250 bn backstop for OpenAI’s 10GW Ohio campus
Nvidia is making moves far beyond chips. Reports surfaced that Nvidia is in talks to provide up to $250 billion in financial guarantees to backstop OpenAI. The massive structure supports OpenAI’s effort to lease a 10-gigawatt data center campus in Piketon, Ohio. Developed by SoftBank’s SB Energy on a former uranium enrichment site, the overall project is estimated to cost upwards of $500 billion.
New AI model debut from Anthropic while Kimi K3 goes open-weight
Anthropic officially made Claude Opus 5 available to developers and enterprise clients. Rather than competing solely on per-token pricing, Anthropic emphasised “cost per accepted task” efficiency, targeting autonomous agentic workflows. Meanwhile, China’s Moonshot AI released open weights for Kimi K3, delivering one of the largest open-weight models available to global developers.
Meta faces ethics scrutiny while xAI sues Minnesota
Meta experienced a turbulent week as its stock fell nearly 8 per cent following weaker-than-expected earnings, despite heavy AI capital investments. Adding to the pressure, reports revealed Meta’s internal “Operation Cannes,” where contractors posed as minors to test rival chatbots (ChatGPT, Gemini, Character.AI) with crisis prompts, sparking intense debates on tech safety ethics.
Meanwhile, Elon Musk’s xAI filed a lawsuit against the state of Minnesota regarding its first-in-the-nation legislation banning “nudification” technology, thus setting up a high-profile legal battle over state-level AI regulations. This might start a new precedent for AI companies defending their cutting-edge AI models and their capabilities.