OpenAI Launches GPT-5.6 — But Only for a Chosen Few
In a move that caught the industry off guard, OpenAI released GPT-5.6 on Thursday — not with the broad rollout the company is known for, but as a limited preview restricted to roughly 20 government-approved organizations. The release comes after the White House requested that OpenAI limit access pending security assessments under the June 2 executive order on AI innovation and security.
The GPT-5.6 series introduces three tiers: Sol, the flagship model with enhanced coding, biology, and cybersecurity capabilities; Terra, a balanced mid-tier option matching GPT-5.5 performance at half the cost; and Luna, a fast, affordable model for high-volume workloads. Sol features a new "ultra" reasoning mode that deploys sub-agents for complex tasks, along with what OpenAI calls its "most robust safety stack to date."
OpenAI was notably candid about its discomfort with the arrangement. "We don't believe this kind of government access process should become the long-term default," the company stated. "It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them." The company said it is working with the administration to establish a repeatable framework for future releases and expects broader availability "in the coming weeks."
Anthropic Accuses Alibaba of Largest Known AI Model Theft
In a letter to the U.S. Senate Committee on Banking, Housing, and Urban Affairs, Anthropic revealed that Alibaba-affiliated operators ran nearly 25,000 fraudulent accounts to extract capabilities from its Claude models in what the company described as "the largest known distillation attack on Anthropic to date."
Between April 22 and June 5, these accounts generated approximately 28.8 million exchanges with Claude, specifically targeting the Mythos Preview model. Distillation is an AI training technique where a smaller model is trained on the outputs of a more capable one — effectively copying its intelligence without building it from scratch. Anthropic tied similar extraction efforts to DeepSeek, Moonshot AI, and MiniMax.
The revelation adds fuel to the intensifying U.S.–China AI rivalry and raises serious questions about the enforceability of model access controls. Alibaba did not respond to requests for comment.
100+ Experts Fight Ban on Anthropic's Mythos and Fable 5
The fallout from Anthropic's Mythos model penetrating classified government systems in hours during Project Glasswing testing continues to ripple through the security community. While the administration restricted both Mythos 5 and Fable 5 to U.S. citizens on June 12 — effectively taking them offline — more than 100 cybersecurity experts have signed an open letter urging reversal of the restrictions.
U.S. officials have since clarified that Mythos "surfaced the weaknesses quickly but did not necessarily exploit them" — an important distinction that the signees say strengthens the case for using such models as defensive tools rather than banning them. The debate highlights a fundamental tension in AI safety: the same capabilities that make a model dangerous for offense may make it invaluable for defense.
GPT-5 Pro Cracks a T-Cell Mystery That Stumped Scientists for Years
In a striking demonstration of AI's growing role in scientific discovery, immunologist Derya Unutmaz at The Jackson Laboratory used OpenAI's GPT-5 Pro to analyze immune-cell data that had stalled his team since 2022. The model surfaced T-cell gene-expression patterns that human researchers had missed for years, compressing months of analytical work into minutes.
Unutmaz emphasized that AI served as a "hypothesis partner" rather than a researcher replacement — generating insights that his team then validated through traditional laboratory methods. The case adds to a growing body of evidence that frontier AI models are becoming genuinely useful scientific instruments, not just impressive demos.
AI Legislation Sweeps Through State Capitols
While Washington debates the Great American AI Act at the federal level, state legislatures are moving fast. In the past week alone:
California sent AB 2148, a bill banning AI public school teachers, to Governor Newsom's desk. The bill requires that only "natural persons" serve as classroom instructors.
New York wrapped its session with seven AI bills awaiting Governor Hochul's signature, including kids' chatbot safety measures, a training data transparency act, and a one-year moratorium on hyperscale data centers.
Rhode Island Governor Dan McKee signed three AI laws covering chatbot self-harm prevention, restrictions on AI-delivered therapy, and mandatory patient notification when healthcare providers use AI documentation tools.
Arizona Governor Katie Hobbs vetoed three AI bills — including a chatbot safety measure and an AI surveillance pricing ban — among 88 bills rejected in a single day, without explanation.
With Colorado's Consumer Protections for AI Act taking effect June 30, the patchwork of state-level AI regulation is growing fast, adding pressure on Congress to establish a unified federal framework.