The AI industry spent Thursday in court, in code reviews, and in a growing argument about who actually owns the models powering your spreadsheet. OpenAI got slapped with a sanctions motion from The New York Times and a coalition of news publishers accusing it of hiding 78 million ChatGPT conversations. Meanwhile, the Ethereum Foundation reported that AI agents found genuine security bugs in its protocol code, and OpenAI’s GPT-5.6 landed inside Microsoft 365 Copilot just days after reports Microsoft is quietly building its own alternatives. All of these stories are connected by the same thread: the gap between what AI companies say and what AI companies do is getting harder to ignore.
OpenAI Faces Sanctions Motion Over 78 Million Hidden ChatGPT Logs
News organizations led by The New York Times asked a federal judge on Thursday to sanction OpenAI for allegedly concealing a 78 million conversation dataset and deleting billions of ChatGPT logs despite a standing preservation order. The motion, filed in Manhattan federal court, also names the New York Daily News, Ziff Davis, and the Center for Investigative Reporting as plaintiffs.
According to the filing, OpenAI spent more than two years claiming it could not search its own training corpus or ChatGPT conversations for copyrighted news content. That position collapsed in April, when a privacy engineer named Vincent Monaco revealed under deposition that OpenAI had already run such searches internally. His testimony reportedly showed the company had amassed roughly 78 million de-identified conversations before the lawsuit began and had screened them for news content. None of this was disclosed while publishers spent months combing through a heavily redacted sample of 20 million logs, which the court later deemed unusable.
The plaintiffs want the court to bar OpenAI from relying on that 20 million log sample at trial, to find that ChatGPT outputs contained substantial reproduction of copyrighted work, and to block any argument to the contrary. They also want attorneys’ fees and a jury instruction about the deleted logs. Lead counsel Ian Crosby said OpenAI “lied to The Times, The Daily News Plaintiffs, the public, and the court.” OpenAI spokesperson Drew Pusateri called the claims blatantly false and said the company will “continue defending our users’ privacy and the long-established principles of fair use.” The Times has spent more than $28 million fighting AI companies in court so far, including a separate suit against Perplexity.
Here’s what this tells us: the discovery phase of AI copyright litigation is turning into a battlefield where internal tooling and engineering practices matter as much as legal arguments. If the court grants these sanctions, it could fundamentally limit OpenAI’s ability to defend itself at trial. The case has been moving since December 2023, and the stakes keep climbing.
Ethereum Turns to AI Agents for Security Testing
The Ethereum Foundation’s Protocol Security team reported that coordinated AI agents found genuine vulnerabilities in Ethereum’s code, including a remotely triggered panic in the peer-to-peer layer used by consensus clients. Developers have already fixed that issue. The foundation divided the workflow into reconnaissance, hunting, gap-filling, and validation roles, drawing on earlier research from Anthropic and Cloudflare.
The agents performed well when linking specifications to code and proposing possible root causes. But the team said the real challenge wasn’t finding bugs. It was telling real bugs from convincing fakes. “The real surprise was how little of the work went into finding them, and how much went into telling the real bugs from the ones that just looked real,” the team wrote. Agents sometimes treated unreachable call chains as exploitable and overstated severity. Human reviewers remain essential for validating findings and judging practical impact.
This matters beyond crypto. The Ethereum experiment is one of the first real-world deployments of AI agents doing security work on production infrastructure, and the results are instructive. Agents expand coverage dramatically but they also produce a much larger pile of confident-sounding false positives. The foundation said agents “let us cover far more ground than we could by hand” but in exchange they “ask for more careful judgment, across a much bigger pile of confident-sounding claims.” That tradeoff is going to show up everywhere AI agents touch security work.
The experiment comes as Ethereum prepares for what Vitalik Buterin has called the Lean Ethereum roadmap, a four-year plan affecting multiple network components while targeting faster performance and quantum-computing resistance. The foundation has cut staffing by 20% and reduced its annual budget by nearly half. The last comparable transition was The Merge, which moved Ethereum from proof-of-work to proof-of-stake validation.
GPT-5.6 Lands in Microsoft 365 as the Partnership Frays
OpenAI announced that GPT-5.6 will become the preferred model in Microsoft 365 Copilot, running across Word, Excel, PowerPoint, Copilot Chat, and Cowork. The flagship Sol variant costs $5 per million input tokens and $30 for output. Terra runs $2.50/$15, and Luna sits at $1/$6. Sam Altman told CNBC that Sol is 54% more token efficient on agentic coding tasks, framing the release around cost as much as capability.
The timing is awkward. Two days before the announcement, it was reported that Microsoft had begun answering some Word and Excel prompts with its own MAI models, part of a broader push to rein in AI costs. Microsoft added seven new MAI models at its Build conference last month. The company has poured roughly $13 billion into OpenAI since 2019 but lost its exclusive cloud provider status in January 2025 and gave up exclusive access to OpenAI’s intellectual property in April. Nothing in the new “preferred model” language rules out Microsoft leaning harder on in-house systems.
The three-way model war is getting clearer. Yellow’s comparison of GPT-5.6, Grok 4.5, and Fable 5 lays out the pricing spread: GPT-5.6 Luna at $1 per million input tokens, Grok 4.5 at $2, and Fable 5 at $10. Fable 5 still leads on the hardest coding tests with 80.4% on SWE-Bench Pro, but GPT-5.6 Sol hits 91.9% on Terminal-Bench 2.1 through a new ultra mode that splits work across parallel subagents. A completed agentic coding task costs $2.49 on Grok 4.5 versus $11.80 for Fable 5 inside Claude Code. That price gap is reshaping purchasing decisions, and analysts increasingly describe a routing market rather than a single winner, with cheap tiers absorbing bulk work while expensive models handle the hardest problems.
OpenAI Raises Bio Bug Bounty to $50,000
OpenAI evolved its GPT-5.5 Bio Bug Bounty into an ongoing private program called the OpenAI Bio Bounty Program, focused on universal jailbreaks that can defeat predefined biosafety challenges against its frontier models. The reward for a universal jailbreak has been raised from $25,000 to $50,000 for both GPT-5.6 and GPT-5.5. Testing for GPT-5.5 ends July 27, after which only GPT-5.6 will be in scope. Accepted applicants need existing ChatGPT accounts and must sign an NDA.
This is a different approach than Anthropic’s classifier-based defense for Fable 5. OpenAI is essentially crowdsourcing red teaming for biosafety, paying outsiders to find the exact techniques that could make their models dangerous. The program runs alongside OpenAI’s existing Safety Bug Bounty and Security Bug Bounty programs on Bugcrowd. The doubling of the reward suggests either that the original $25,000 wasn’t generating enough serious submissions, or that GPT-5.6’s capabilities have raised the stakes enough to justify higher payouts. Either way, it signals that biosafety jailbreaks remain an unsolved problem.
Quick Hits
OpenAI also published a detailed audit of SWE-Bench Pro, finding that roughly 30% of the 731-task dataset is broken. The problems fall into four buckets: overly strict tests, underspecified prompts, low-coverage tests, and misleading prompts. OpenAI’s team used Codex-based investigator agents plus five experienced engineers per task to flag issues. The company previously abandoned SWE-Bench Verified for similar contamination and design problems. This is honesty about benchmark quality that the industry needs more of.
Hugging Face published Part 3 of its PyTorch profiling series, focused on attention mechanisms. The post walks through naive attention, in-place causal masking, SDPA, and custom kernels under the profiler, showing exactly which operations generate unexpected memory copies and how in-place operations can eliminate them. Practical engineering content for anyone optimizing transformer inference.
OpenAI launched GPT-Live, a full-duplex voice model that can listen and speak simultaneously. It delegates complex reasoning to GPT-5.5 in the background while keeping the conversation flowing. The model handles interruptions, pauses, and background noise better than Advanced Voice Mode, and 150 million people use ChatGPT voice features weekly. This is the most natural voice AI has felt so far, though the background model is still GPT-5.5, not GPT-5.6.
Anthropic proposed an industry-wide framework for scoring jailbreak severity, developed with Amazon, Microsoft, Google, and other Glasswing partners. The framework emerged from the Fable 5 export control saga, where an Amazon-reported jailbreak triggered government intervention. Anthropic trained a new safety classifier that blocks the reported technique in over 99% of cases, at the cost of more false positives during routine coding. This is the industry trying to self-regulate before governments step in harder.
Mistral shipped Robostral Navigate, an 8B model that handles robot navigation using a single RGB camera with no depth sensors or LiDAR. It hits 76.6% on R2R-CE validation unseen, beating multi-sensor approaches. Trained entirely in simulation on 400,000 trajectories across 6,000 scenes, with a prefix-caching trick that compresses training tokens by 22x. Reinforcement learning added another 3.2% success rate on top of supervised training. The model generalizes across wheeled, legged, and flying robots.
NVIDIA published a deep dive on open data for agents as part of its Nemotron release, arguing that synthetic data is how the industry scales agent training. The post covers the Nemotron Post-Training v3 Prompt Atlas, Nemotron-Personas (locally grounded synthetic personas representing 2.4 billion people across 10 countries), and the tradeoffs between synthetic and real data. NVIDIA also announced that Nemotron 3 Ultra achieves benchmark-leading performance with LangChain’s Deep Agents harness at 10x lower inference cost than closed models, with no model retraining required. Every gain came from engineering the harness around the model.
Rundown for July 11, 2026. Sources: Yellow, OpenAI, Anthropic, Mistral AI, NVIDIA, Hugging Face.