AI INTELLIGENCE / DAILY SIGNAL
0011011011
2026-06-26 / AI NEWS

AI News Intelligence Digest — June 26, 2026

Anthropic accuses Alibaba of largest-ever AI distillation attack. Between April 22 and June 5, Alibaba's Qwen lab ran 28.8 million fraudulent API exchanges across 25,000 fake accounts targeting Claude's software engineer...

BRIEFING STRUCTURE22 SIGNALS

Inside this digest

  • Top Stories
  • Model Updates
  • Open Source
  • Industry Moves
  • Safety & Policy
  • Worth Watching
← Back to archive

Top Stories

  • Anthropic accuses Alibaba of largest-ever AI distillation attack. Between April 22 and June 5, Alibaba's Qwen lab ran 28.8 million fraudulent API exchanges across 25,000 fake accounts targeting Claude's software engineering and agentic reasoning capabilities. Anthropic took the evidence to the Senate Banking Committee and the White House, framing model distillation as a national security threat. Senators Hagerty and Kim are advancing an amendment to defense legislation to sanction firms caught stealing AI model outputs. This follows a February incident where DeepSeek, Moonshot, and MiniMax ran 16M queries across 24K accounts — the campaigns are doubling in scale. (CNBC — "Anthropic accuses Alibaba of campaign to 'brazenly' extract AI capabilities"; Decrypt — "Anthropic Urges Congress to Crack Down on AI Distillation By Chinese Rivals")

  • OpenAI and Broadcom unveil Jalapeño — OpenAI's first custom chip. Designed from scratch in nine months (believed to be the fastest high-performance ASIC development cycle ever), Jalapeño targets ~50% lower inference cost per token vs Nvidia GPUs. OpenAI used its own models to accelerate chip design. TSMC manufactures; initial deployment late 2026, full scale H1 2028. Broadcom is now the neutral silicon partner designing custom chips for Google (TPUs), OpenAI, Meta (MTIA), and ByteDance. (OpenAI blog; TechCrunch — "OpenAI unveils its first custom chip, built by Broadcom")

  • Google DeepMind exodus: four senior departures in six days. Noam Shazeer (co-author "Attention Is All You Need") left for OpenAI June 18. Nobel laureate John Jumper (AlphaFold), Jonas Adler, and Alexander Pritzel all heading to Anthropic June 20-24. Internal GPU reallocation politics — compute pulled from active projects — cited as the primary frustration. GOOG down ~6% in five sessions. (Bloomberg; Fortune — "As top talent leaves Google DeepMind, some question if the lab can remain at the forefront")

  • GPT-5.6 slips to July. Polymarket odds for a June 22-28 launch collapsed from ~83% to ~18%. July release now priced at 94%. Expected: 1.5M token context window (43% increase over GPT-5.5), improved agentic task completion, sharper coding. Internal codename "kindle-alpha." OpenAI's ~6-week release cadence continues. (Yahoo Finance — "Traders abandon bets on a GPT-5.6 launch this week")

  • Fable 5 & Mythos 5 still dark — day 14. Commerce Department export ban since June 12 remains in effect. Anthropic staff confirmed zero Fable/Mythos traffic is being served. Anthropic argues the cited jailbreak also works on GPT-5.5 (which faces no ban) and calls the recall disproportionate. No restoration timeline. (Anthropic — "Statement on the US government directive"; Fortune)

Model Updates

  • FrontierCode benchmark launched by Cognition — production-quality eval focusing on scope control, regression safety, and test quality. Fable 5 scored 46.3%, well ahead of Opus 4.8 (34.3%) and GPT-5.5 (25.5%). Validates Fable 5's coding lead — if only it were available.
  • Gemini 3.5 Pro delayed to July. Google missed its June GA target. Gemini 2.5 Pro with Deep Think still strong (82.4% GPQA Diamond). Gemini 3.5 Live Translate shipped for Pixel 10/Galaxy S26 — speech-to-speech in 70+ languages.
  • Apple AFM 3 foundation models announced — five models using Instruction-Following Pruning (1-4B active params from 20B base). AFM 3 Cloud Pro runs on Nvidia GPUs in Google Cloud. Apple's on-device AI play is real.
  • DeepSeek V4-Pro 75% price cut now permanent — $0.44/M input, $0.87/M output. Undercuts GPT-5.5 by 5-17x. The pricing pressure on Western labs is relentless.
  • AI coding market hits $9.3B (Mordor Intelligence). Claude Code leads at ~40% market share. Growing 26% annually to ~$30B by 2031.

Open Source

  • NVIDIA Cosmos 3 — first open omnimodel for physical AI. Trained on 20T tokens of multimodal data (1B images, 400M videos, audio, robot actions). Three variants: Super (highest physics accuracy), Nano (fast inference), Edge (coming). Open under Linux Foundation's OpenMDW-1.1 license. Rankings: #1 on Physics-IQ, PAI-Bench, RoboLab, and RoboArena.
  • MiniMax M3 — first open-weight model combining frontier SWE capability (59.0% SWE-Bench Pro), 1M token context, and native computer use.
  • Zyphra ZAYA1-8B — Apache 2.0, sparse routing with only 760M active params per token out of 8B total. Efficiency frontier.

Industry Moves

  • Anthropic nearing first operating profit — tracking ~$559M in Q2 2026. Still valued at $965B (Series H), above OpenAI's $852B.
  • MGX (UAE) raised $50B for AI infrastructure — one of the largest single funding closes ever. Regional sovereign wealth + global pension funds.
  • Qualcomm acquiring Modular for ~$3.9B — diversifying beyond smartphones into data-center AI. Modular's platform runs AI across multiple chip types without code rewriting.
  • Frontier AI pricing surged across the board. GPT-5.5 doubled GPT-5.4's pricing. Claude Fable 5 launched at $10/$50 per M tokens (2x Opus 4.8). OpenAI spent $1.35 per revenue dollar in 2025 — cheap AI is over.

Safety & Policy

  • Trump signed "Promoting Advanced AI Innovation and Security" EO (June 2). Establishes voluntary 30-day pre-release government review for frontier models (cut from 90 days in earlier drafts). Framework due by August 1. Classified benchmarking process for cyber capabilities in development.
  • EU AI Act high-risk deadlines delayed. European Parliament approved amendments June 16: standalone high-risk AI (Annex III — education, employment, law enforcement) pushed from August 2026 to December 2027 (16-month delay). Product-embedded AI delayed to August 2028. Industry pressure won the reprieve.
  • Colorado repealed and replaced SB 24-205 with narrower SB 26-189 (automated decision-making tech), effective January 2027. California's SB 53 (Transparency in Frontier AI Act) requires risk frameworks, safety incident reporting, and whistleblower protections from frontier model developers.

Worth Watching

  • "AI Daily Brief" by Nathaniel Whittemore — consistent daily coverage of the Anthropic/Alibaba distillation story, DeepMind departures, and GPT-5.6 timeline shifts. Best daily podcast-style AI news. https://www.youtube.com/@AIDailyBrief
  • Broadcom/OpenAI Jalapeño deep-dive — Tom's Hardware published the most technically detailed breakdown of the chip: reticle-sized ASIC, 9-month development cycle, co-design with OpenAI models. Search "Broadcom OpenAI Jalapeño" on YouTube for video analysis.

Developing stories from yesterday's digest: DeepMind brain drain has accelerated (now 4 departures, up from 2). Fable/Mythos ban enters day 14 with no movement. GPT-5.6 launch window has slipped from June to July. SpaceX Colossus compute deals continue to accumulate. Anthropic distillation allegations are NEW and represent a significant escalation from the February DeepSeek incident.