This Week in AI: June 20–26, 2026
An IP-theft accusation against Alibaba escalates US-China AI tension, a red-team test shows how deep frontier models can probe classified systems, and labs race to lock down chips, memory, and top scientific talent.
On this page+
- The Big Story: Anthropic Accuses Alibaba of Largest-Ever AI Distillation Attack
- Source
- Anthropic's Mythos Model Breaks Into Classified US Systems During Authorized Red-Team Test
- Source
- Nobel Laureate John Jumper Leaves Google DeepMind for Anthropic
- Source
- Google's Dataland Becomes the World's First AI Arts Museum
- Source
- OpenAI's GPT-5.5-Cyber Outscores Anthropic's Banned Mythos on Security Benchmarks
- Source
- Micron and Anthropic Announce Strategic AI Infrastructure Agreement
- Source
- OpenAI and Broadcom Unveil Jalapeño, a Custom AI Inference Chip
- Source
- Strategic Takeaway
AI competition entered a sharper, more adversarial phase this week. An IP-theft accusation against Alibaba escalated US-China tension from export controls into outright capability theft, a red-team exercise revealed how deeply frontier models can probe classified government systems, and labs raced to lock down chips, memory supply, and marquee scientific talent. Cultural deployment kept pace too, with AI's first dedicated museum opening its doors in Los Angeles.
The Big Story: Anthropic Accuses Alibaba of Largest-Ever AI Distillation Attack
In a letter to the Senate Banking Committee dated June 10 and made public this week, Anthropic accused Alibaba-linked operators of running the largest known distillation attack against its models to date: roughly 28.8 million exchanges through about 25,000 fraudulent accounts between April 22 and June 5, allegedly used to extract Claude's capabilities into rival models. The accusation lands as Anthropic's own Fable 5 and Mythos 5 remain suspended under a US government export control order, sharpening the US-China divide over who gets to build on frontier AI.
Source
Anthropic's Mythos Model Breaks Into Classified US Systems During Authorized Red-Team Test
Senator Mark Warner told a congressional hearing that NSA chief Joshua Rudd informed him Anthropic's Mythos model "broke into almost all" of the agency's classified systems "not in weeks, but in hours" during Project Glasswing, a restricted exercise where intelligence agencies ran Mythos against their own environments. Officials were careful to clarify this was a self-directed red-team test, not an external intrusion, and that finding a vulnerability is not the same as exploiting one — but the result still reframes how national security agencies are weighing frontier models as both an asset and a risk.
Source
Nobel Laureate John Jumper Leaves Google DeepMind for Anthropic
AlphaFold co-creator and 2024 Nobel Chemistry laureate John Jumper announced June 20 that he is leaving Google DeepMind after nearly nine years to join Anthropic, a move that lands the same week fellow DeepMind veteran Noam Shazeer departed for OpenAI. Neither side has confirmed Jumper's exact role, but the hire fits Anthropic's growing push into computational biology and life sciences.
Source
Google's Dataland Becomes the World's First AI Arts Museum
Dataland, a 25,000-square-foot museum co-founded by media artist Refik Anadol and Efsun Erkılıç with Google as technology and creative collaborator, opened to the public June 20 at The Grand LA in downtown Los Angeles. Its inaugural exhibition, "Machine Dreams: Rainforest," runs on Google's Large Nature Model and generates roughly 1.2 billion pixels of generative imagery in real time; the museum will also fund a $25,000 artist-residency grant for four artists per six-month cohort.
Source
OpenAI's GPT-5.5-Cyber Outscores Anthropic's Banned Mythos on Security Benchmarks
OpenAI released a more capable GPT-5.5-Cyber on June 22 as part of its "Patch the Planet" defensive-security push, posting a record 85.6% on the CyberGym benchmark with notable gains on exploit-generation and long-horizon vulnerability discovery tests. The model stays restricted to vetted defenders through OpenAI's Daybreak Cyber Partner Program rather than shipping via public API.
Source
Micron and Anthropic Announce Strategic AI Infrastructure Agreement
Micron Technology and Anthropic announced a strategic agreement on June 22 spanning joint memory and storage architecture design for next-generation AI infrastructure, alongside a Micron investment in Anthropic's Series H round. The deal gives Anthropic a dedicated hardware partner as compute and memory supply increasingly become competitive bottlenecks for frontier labs.
Source
OpenAI and Broadcom Unveil Jalapeño, a Custom AI Inference Chip
OpenAI and Broadcom unveiled Jalapeño on June 25, OpenAI's first custom "Intelligence Processor" built specifically to accelerate LLM inference, taken from initial design to manufacturing tape-out in just nine months. Initial deployment is targeted for the end of 2026, joining a growing list of frontier labs designing their own silicon rather than relying solely on third-party GPUs.
Source
Strategic Takeaway
This week shows frontier AI competition fracturing along several fronts at once: an IP-theft accusation against Alibaba pushes US-China tension from export controls into outright capability theft, while Mythos's red-team results against classified systems and OpenAI's defenders-only GPT-5.5-Cyber show frontier labs operating more like national-security infrastructure than ordinary software vendors. Meanwhile, the Micron and Broadcom deals reveal the real contest increasingly runs through who controls chip and memory supply chains, and a wave of marquee scientist moves between labs shows talent consolidating around whichever lab looks best positioned to win that contest, even as AI's reach keeps expanding into unexpected cultural spaces like Dataland.
