Skip to content
Daily Brief News Daily Brief

13 AI stories from June 26, 2026: GPT-5.6 federal gating, Alibaba distillation attack, IBM sub-1nm chip, OpenAI Jalapeno, AI data center moratorium, GLM-5.2, Fable 5 day 14, EU GPAI fines in 37 days, Qualcomm buys Modular, Gemini 3.5 Flash computer use, Google loses five researchers, fake agent skill reaches 26,000 agents, and GPT-4.5 retires tomorrow

· by Pondero Newsdesk · 13 stories

AI news daily brief: 2026-06-26

Thirteen stories today: federal approval now gates a US AI model before launch for the first time, Anthropic presented the Senate with hard numbers on a Chinese distillation campaign, IBM and OpenAI each revealed chip architectures that alter the cost curve for AI inference, and a bipartisan bill would freeze large data center construction entirely until Congress acts.

US government gates GPT-5.6 behind federal approval in the first pre-launch model review of any American AI system

The Trump administration formally asked OpenAI to release GPT-5.6 only to a curated set of government-approved enterprise partners rather than the public, per TechCrunch. The Office of the National Cyber Director and the Office of Science and Technology Policy made the request as the administration builds a testing and evaluation framework for frontier models. OpenAI CEO Sam Altman confirmed the arrangement in an internal Q&A, calling it "limited enterprise access" with the government approving customers individually. Altman described the cooperation as voluntary but "not sustainable long term." GPT-5.6 is on par with Anthropic's Mythos 5 in advanced cybersecurity capabilities, per government briefings reported by CybersecurityNews. No public launch date was given. This marks the first time a US government office has formally reviewed and gated a domestic AI model before public release.

Full story: US government gates GPT-5.6 behind federal approval

Anthropic tells Senate that Alibaba ran 25,000 fake accounts and 28.8 million Claude queries in the largest AI distillation attack ever disclosed

Anthropic formally accused Alibaba of running "the largest known distillation attack on Anthropic to date" in a letter to the US Senate Banking, Housing, and Urban Affairs Committee, per CNBC. The campaign ran April 22 to June 5, 2026. Per Anthropic, 24,975 fraudulent accounts using commercial proxy services ran 28.8 million exchanges with Claude models to extract software-engineering and agentic-reasoning capabilities. Anthropic head of policy Sarah Heck told senators the attack was "carried out illicitly, systematically, and at industrial scale to harvest U.S. AI capabilities across frontier labs and repackage them as their own." Alibaba had not publicly responded as of June 26. Anthropic disclosed three smaller campaigns in February 2026 involving DeepSeek, Moonshot, and MiniMax.

Full story: Alibaba ran 28.8M Claude queries in distillation attack, Anthropic tells Senate

IBM unveils sub-1nm nanostack chip with 100 billion transistors and a projected 50% performance gain over its 2nm node

IBM Research announced on June 25 the world's first sub-1 nanometer chip technology using a "nanostack" architecture operating at the 0.7nm (7-angstrom) node, per the IBM Newsroom. The nanostack approach vertically stacks and staggers nanosheet transistors using 3D sequential integration. The result is nearly 100 billion transistors on a fingernail-sized chip, roughly double the density of IBM's current 2nm node. IBM projects either a 50% performance increase or a 70% energy efficiency improvement over 2nm. Production-ready chips are "about five years away," per IBM. The primary target workloads are AI inference and high-bandwidth compute.

Full story: IBM sub-1nm nanostack chip doubles transistor density over 2nm

OpenAI and Broadcom launch Jalapeno, a custom inference chip targeting 50% lower AI serving cost than Nvidia GPUs

OpenAI and Broadcom jointly announced "Jalapeno" on June 24: a custom LLM-optimized inference chip built from scratch in nine months using OpenAI's own AI models to accelerate chip design, per OpenAI's announcement page. The processor targets roughly 50% cost reduction versus current Nvidia GPU workloads, per the announcement. Initial deployment targets gigawatt scale by end of 2026 with data center partners including Microsoft. TechCrunch called it OpenAI's first custom chip. Jalapeno is a direct response to the company's dependence on Nvidia for ChatGPT and API serving capacity.

Full story: OpenAI and Broadcom launch Jalapeno chip to cut AI serving cost 50%

Sanders and Ocasio-Cortez introduce AI Data Center Moratorium Act to freeze construction above 20MW until Congress acts

Representative Alexandria Ocasio-Cortez introduced the House version of the AI Data Center Moratorium Act on June 25, mirroring the Senate bill by Senator Bernie Sanders, per the Sanders press release. The legislation would impose a nationwide halt on constructing or expanding AI data centers with power demand of 20 megawatts or more until Congress passes "strong national safeguards." Nine House members co-sponsored the bill. The legislation directly conflicts with the Trump administration's FERC fast-track order from June 18 and the White House executive order on AI innovation. Operators planning data center buildouts above the 20MW threshold should monitor the bill's committee progress.

Full story: Sanders and Ocasio-Cortez introduce AI Data Center Moratorium Act

GLM-5.2 from Zhipu AI beats GPT-5.5 on SWE-bench Pro at one-sixth the cost with MIT-licensed open weights

Zhipu AI (rebranding as Z.ai) released GLM-5.2 on June 13 and published open weights on June 16 under the MIT license. The model uses a 744B-parameter Mixture-of-Experts architecture with 40 billion active parameters per token and a 1-million-token context window. On SWE-bench Pro, GLM-5.2 scores 62.1, ahead of GPT-5.5 at 58.6, while trailing Claude Opus 4.8 at 69.2, per VentureBeat. API pricing is roughly one-sixth the cost of GPT-5.5. Latent Space (Swyx) called it the first open model to genuinely pass the "frontier model" vibe check. Z.ai said it will release "Open Fable," an open-weight model at Mythos-class capability, by December 2026.

Full story: GLM-5.2 beats GPT-5.5 on coding benchmarks at one-sixth the cost

Fable 5 ban day 14: Lutnick deadline passes without public response as Tom Brown leads White House talks

As of June 26, day 14 of the US government export-control directive, Claude Fable 5 and Claude Mythos 5 remain offline globally, per Anthropic's statement page. Commerce Secretary Lutnick faced a June 26 deadline from four bipartisan House members to explain the ban in writing; no public response had been issued by publication time. WIRED and Gizmodo reported June 24 that Anthropic co-founder Tom Brown has replaced CEO Dario Amodei in White House negotiations, per Gizmodo. Trump said at the G7 on June 19 that Anthropic is "no longer a national-security threat." Anthropic's July 8 biometric ID policy is the structural mechanism most analysts cite for a US-citizens-only partial restoration. Operators should hold fallback models in production until a date is confirmed.

EU AI Act GPAI enforcement starts August 2 with fines up to 3% of global revenue for OpenAI, Google, Anthropic, and Mistral

The EU AI Act's enforcement powers for General Purpose AI model providers enter full application under Article 101 on August 2, 2026 (37 days away), per Latham and Watkins. From that date, the European AI Office can fine providers up to 3% of global annual turnover or EUR 15 million for violations of GPAI transparency, copyright, and systemic-risk management obligations. The GPAI Code of Practice (published July 2025) is a voluntary compliance tool that mitigates but does not eliminate fine risk. OpenAI, Google, Anthropic, and Mistral are all subject to the obligations where their models are available in the EU. The minimum action before August 2 is confirming that transparency documentation and systemic-risk assessments are filed with the AI Office.

Qualcomm agrees to buy Modular for $4 billion to challenge Nvidia's CUDA stack with Chris Lattner's MAX AI Engine

Qualcomm announced on June 24 an agreement to acquire Modular Inc. for approximately $4 billion in an all-stock deal, per the Qualcomm investor release. Modular was founded by Chris Lattner, creator of LLVM, MLIR, and Swift, alongside Tim Davis. Modular's flagship product is the MAX AI Engine, a hardware-portable inference framework that runs AI models across CPU, GPU, and custom silicon without architecture-specific rewrites. Qualcomm will integrate the stack into its data center product line. The deal is expected to close in H2 2026 pending regulatory clearance. Modular raised $250 million at a $1.6 billion valuation in September 2025, per CNBC, putting the acquisition at a 2.5x premium nine months later.

Google ships Gemini 3.5 Flash with native computer use, 1M-token context, and 4x throughput in a single agent loop

Google announced on June 24-25 that computer use is natively integrated into Gemini 3.5 Flash at general availability, per the Google AI Blog. An agent using Flash can see a screen, click, type, and navigate web browsers and desktop apps directly, alongside Google's built-in Search and Maps grounding tools, all in a single unified agent loop. Flash supports a 1-million-token context window and runs at roughly 4x the throughput of Gemini 3.1 Pro. The model is available via Gemini API, Vertex AI, and the Gemini Enterprise Agent Platform. Gemini 3.5 Pro with Deep Think reasoning remains in limited enterprise preview only. For teams evaluating computer-use agents in production, Flash is the first Google model where the capability is available without a preview waitlist.

Google loses five senior AI researchers in eight days as two more Gemini leads move to Anthropic

Bloomberg reported June 24 that Jonas Adler and Alexander Pritzel, both key contributors to Google DeepMind's Gemini model line, are leaving to join Anthropic, per Bloomberg. The two follow John Jumper (Nobel Chemistry laureate for AlphaFold, joined Anthropic June 20), Arthur Conmy (Gemini post-training and mechanistic interpretability, joined Anthropic June 25), and Noam Shazeer (co-author of the 2017 Transformer paper, joined OpenAI June 18). Five named departures landed in eight days. Alphabet stock dropped 7.2% this week, per TechCrunch. The pattern matters for teams building on Gemini: a sustained talent drain at the lead model team raises questions about roadmap continuity past 3.5 Pro.

A fake AI agent skill cleared every security scanner and installed on 26,000 agents before responsible disclosure

Security research firm AIR built a fake AI agent skill named "brand-landingpage," submitted it to a popular open-source skill repository, and documented it installing on approximately 26,000 agents, including corporate accounts, before public disclosure on June 23, per The Hacker News. Every AI skill security scanner AIR tested cleared the skill as safe. The vulnerability: the skill used a mutable external link whose behavior could be changed after the scanner's initial approval. No harmful payload was deployed; this was a responsible disclosure proof-of-concept. The research identifies a class of supply-chain risk present in any skill marketplace that scans only at submission time rather than at install or runtime. Teams operating agent fleets should audit whether their skill sources pin to immutable versions.

OpenAI retires GPT-4.5 from ChatGPT tomorrow, closing the GPT-4 era in the consumer product

OpenAI will retire GPT-4.5 from ChatGPT on June 27, 2026, per the ChatGPT Release Notes. This ends the GPT-4 era in the consumer product: GPT-4.5 was the last GPT-4-family model available to paid ChatGPT subscribers. Paid subscribers who had GPT-4.5 selected will be moved automatically to a default GPT-5.x model. GPT-4.5 was retired from the OpenAI API in July 2025; this is a ChatGPT-product-only change. The removal frees serving infrastructure for GPT-5.x workloads ahead of the expected GPT-5.6 limited enterprise preview. Any ChatGPT-dependent automations that explicitly target GPT-4.5 stop working tomorrow.

Sources