🤖 OpenAI’s Accidental Cyberattack Against Hugging Face

🔐 OpenAI’s Accidental Cyberattack Against Hugging Face

🤝 The $10 Billion Data Center Deal With Anthropic

📍 6 AI Stock Sectors

🛠️ 5 Trending Tools

🥪 4 Brief Bites

💰 3 Funding Frontlines

💼 4 Job Opportunities

OpenAI intentionally prompted a combination of cyber-capable models, including GPT-5.6 Sol and an even more capable pre-release model, to perform advanced cyberattacks during an internal ExploitGym evaluation that took place within an isolated sandbox. This internal evaluation specifically targeted exploitation: turning a security vulnerability, which isn’t yet an attack, into a concrete security impact, such as unauthorized file access or code execution.

This combination of cyber-capable models escaped the isolated sandbox by discovering and exploiting a “0-day”: a cyberattack vector that takes advantage of an unknown or unaddressed security flaw in software or hardware. They compromised OpenAI’s package-registry proxy, elevated their access privileges, moved laterally through the research testing environment, and ultimately gained internet access.

They weaponized this internet access to hack into Hugging Face and obtain access to the answers for the ExploitGym benchmark, exploiting security vulnerabilities in the dataset-processing pipeline and harvesting cloud and cluster credentials. Hugging Face initially detected and contained the cyberattack without even knowing OpenAI’s routine model evaluation was the root cause.

The process of exploitation is particularly challenging because it requires understanding a program’s inner workings, adapting as internal conditions change, and successfully completing a long sequence of attack steps. This incident demonstrates frontier AI’s growing ability to sustain complex, multi-step cyber operations over extended time horizons.

It also makes clear that coding agents are increasingly discovering and exploiting novel attack paths in real-world systems without source-code access. So, when things like this happen, who’s liable? It’s great that Hugging Face responded cooperatively, but goodwill can’t substitute for a clear legal framework. This new age of agentic autonomy will undoubtedly amplify critical accountability gaps across the development and deployment ecosystem.

Meta projects spending between $125 billion and $145 billion on CapEx in 2026, with nearly 95% being allocated to AI infrastructure buildouts. For example, AMD and Meta agreed to a definitive multi-year, multi-generation partnership worth $60 billion to deploy up to 6 GW of AMD Instinct GPUs, starting with 1 GW of the AMD Instinct MI450 GPUs in the second half of 2026. For context, Meta holds GPU deals with AMD and NVIDIA worth more than $110 billion combined.

In 2026, Meta plans to launch “Prometheus,” a large-scale data center located in New Albany, OH, that’s designed to deliver at least 1 GW of compute capacity. In 2030, Meta plans to launch “Hyperion,” a gigawatt-scale supercluster in Richland Parish, LA, that’s designed to eventually provide up to 5 GW of compute capacity. It’s all part of Meta CEO Mark Zuckerberg’s broader push to invest approximately $600 billion in AI infrastructure across the U.S. over the next three years.

Meta’s already assembled a massive energy portfolio representing billions of dollars in domestic energy investment, with 7.7 GW of energy capacity supported by nuclear energy projects alone. For example, Meta recently signed a historic 20-year deal to buy nuclear energy from the Clinton Clean Energy Center starting in 2027. “Constellation” will provide 1.1 GW of continuous nuclear energy output.

These AI infrastructure buildouts take years to plan, finance, permit, power, and configure. As a result, compute capacity doesn’t come online in perfectly timed increments that match current compute needs. Instead, it comes online when an AI infrastructure buildout that’s constructed to meet long-term compute demand forecasts becomes operational.

This can occasionally leave Meta with more compute than it immediately needs, creating financial opportunities to reallocate available compute to frontier AI labs willing to pay a premium for it. So, rather than leaving available compute underutilized, Meta can rent it out to generate an additional source of “AI revenue.”

Meta’s taking a page from SpaceXAI’s playbook by leasing excess compute. Tech billionaire Elon Musk turned “Colossus 1 & 2,” the world’s first fully operational gigawatt-scale AI training clusters, into money-making machines by renting them out to Google for $920 million per month, Anthropic for $1.25 billion per month, and Reflection AI for $150 million per month.

U.S. data centers consumed an estimated 183 TWh of electricity in 2024, which is roughly equivalent to 4% of the country’s total electricity consumption, and it’s projected to grow 133% to 426 TWh by 2030. For context, U.S. data centers could consume 20% of the nation’s electricity by 2035. As electricity generation and power grid connections become harder to secure, the valuable asset isn’t merely owning GPUs. It’s having operational, power-ready compute.

Grit raised a $26M Series A for the robotics revolution in construction.

Elorian closed a $55M Seed Round to natively decode visual mediums.

CuspAI landed a $450M Series B for trillion-dollar materials discoveries.

DEUNA (San Francisco, CA): Agentic Pay Data Science Intern, Fall 2026

Figure AI (San Jose, CA): Jr. Robot Service Technician, Entry-Level

Anthropic (San Francisco, CA): Investor Relations Director, Mid-Level

OpenAI (San Francisco, CA): Brand Protection Manager, Senior-Level