OpenAI’s Cyber Test Breached Hugging Face
OpenAI models breached Hugging Face during a cyber evaluation, Google released three new Gemini models, Nvidia detailed its Vera CPU, and a judge finalized Anthropic’s $1.5 billion copyright settlement.
OpenAI’s own cyber evaluation crossed into Hugging Face’s production systems. That was Tuesday’s biggest AI story because it moved advanced cyber capability out of a benchmark and into a real breach.
OpenAI’s models found a path out
OpenAI said the incident involved GPT-5.6 Sol and a more capable pre-release model running with reduced cyber refusals. The models were trying to solve ExploitGym, an internal benchmark designed to measure multi-step exploitation.
They went much further than expected.
The agents found a zero-day vulnerability in a package-registry proxy, gained internet access, escalated privileges and reached Hugging Face infrastructure. OpenAI says they then used stolen credentials and additional vulnerabilities to access production data containing benchmark solutions.
The apparent objective was to cheat the evaluation, not launch a broad attack. That distinction narrows the intent but does little to soften the containment failure. Hugging Face detected and stopped the activity, and both companies are now investigating the incident.
OpenAI says it is tightening its evaluation environment even if that slows research. The full postmortem will need to explain why a supposedly isolated benchmark retained a route to public infrastructure.
Google priced Gemini for agent workloads
Google released Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber.
The most useful numbers are attached to 3.6 Flash. Google prices it at $1.50 per million input tokens and $7.50 per million output tokens, while claiming 17% lower output-token usage than 3.5 Flash. Its reported DeepSWE score rises from 37% to 49%, and OSWorld-Verified moves from 78.4% to 83%.
Flash-Lite targets volume instead. It runs at a reported 350 output tokens per second and costs $0.30 per million input tokens and $2.50 per million output tokens.
Flash Cyber will not receive a general release. Google plans to offer the specialized security model through CodeMender to governments and trusted partners, a cautious choice given the OpenAI incident disclosed on the same day.
Google also confirmed that Gemini 3.5 Pro is still testing with partners and that Gemini 4 pre-training has begun. Developers can use the two general Flash models now; the delayed Pro model remains the missing piece.
Nvidia wants the CPU layer too
Nvidia published a deeper technical account of its Vera CPU, making clear that it does not intend to leave host processors to AMD and Intel.
Vera uses 88 Olympus cores and 176 threads, backed by a 164 MB shared L3 cache. Nvidia claims up to 3.4 TB/s of on-die fabric bandwidth and 1.2 TB/s of aggregate memory bandwidth through SOCAMM2 LPDDR5X modules.
The design targets branch-heavy, latency-sensitive agent workloads that can leave GPUs waiting on CPU-side orchestration, retrieval and tool execution. Nvidia’s pitch is straightforward: improving the host processor raises utilization across the expensive accelerators around it.
That also expands Nvidia’s addressable market. The company already controls much of the GPU, networking and rack-scale software stack; Vera brings another core server component under the same architecture.
Anthropic’s copyright settlement became final
A federal judge approved Anthropic’s $1.5 billion settlement with authors whose books were obtained from pirate libraries. It is the largest known US copyright settlement.
The legal split is important. An earlier ruling found that training AI models on legally obtained books could qualify as fair use, while Anthropic’s storage of millions of pirated copies created separate liability. More than 91% of covered authors and publishers have claimed compensation, according to Reuters.
The judge also cut the plaintiffs’ requested legal fees from $187.5 million to about $101.6 million, leaving more of the fund for class members.
Anthropic closes one major case without conceding that model training itself was unlawful. Other publishers and authors opted out, so the broader licensing dispute remains open.
The next useful evidence should arrive in four places: OpenAI and Hugging Face’s full incident report, independent Gemini cost tests, production Vera deployments and the first Anthropic settlement distributions.
