Friday, June 19, 2026

Anthropic’s Model Block May Crack

Anthropic’s Model Block May Crack

Today’s Overview

Good morning, AI’s biggest labs are making some sharp turns. Anthropic is trying to bring blocked frontier models back online, Midjourney is suddenly pitching a body scanner and spa business, and agent research is moving from code into robots, security, and privacy. Let’s dive in.

Top Stories

Anthropic Says Mythos and Fable 5 Access Could Return Soon

Anthropic executives said they expect to restore access to Claude Mythos and Fable 5 within days as they work with regulators. The company is framing the disruption as a temporary access issue rather than a permanent product shift.

  • The shutdown followed a White House directive that pushed Anthropic to block the models inside and outside the United States.
  • The Seoul event was meant to focus on Anthropic’s Korea expansion but questions about export controls and Project Glasswing dominated the discussion.
  • Project Glasswing has roughly 150 partners including Google, Nvidia, Microsoft, Apple, Samsung Electronics, SK hynix, and SK Telecom.

Midjourney Is Building a Full-Body Ultrasound Scanner

Midjourney Medical is developing a full-body ultrasound device built around 40 Ultrasound-on-Chip modules and a water-filled ring. The prototype currently moves massive raw data streams to cloud compute and takes 20 minutes because of bandwidth limits, with a stated goal of a 60-second full-body map. Midjourney is also planning a 25,000-square-foot San Francisco spa by late 2027 while it pursues FDA approval.

  • Butterfly Network says each system uses 40 imaging modules under a co-development agreement tied to the scanner.
  • The current sensor ring includes about 358,400 transducers that emit and receive ultrasound as a person descends through water.
  • Midjourney plans to start with body composition maps before broader medical uses that would require additional regulatory clearance.

OpenAI Preps GPT-5.6 for Release

OpenAI is said to be preparing GPT-5.6 for launch next week, potentially with Mini and Pro variants. The reported upgrade centers on a 1.5 million token context window, stronger long-horizon coding, faster Codex responses, and pricing meant to compete aggressively with Anthropic.

  • The rollout may land on Tuesday with the standard model and possible Mini and Pro variants arriving together.
  • Early traces of GPT-5.6 Pro have reportedly appeared for some Pro subscribers.
  • A next-generation voice model called GPT-Bidi-1 is also expected, with bidirectional audio designed to listen and speak at once.

Research & Analysis

Anthropic Tests Claude on Physical Agent Work

Anthropic’s Project Fetch phase two tested software agents on a robotic quadruped, with Claude Opus 4.7 completing hardware programming and sensor integration in under 10 minutes. Anthropic says that was nearly 19 times faster than a Claude-assisted human team and 37 times faster than humans alone, though the system still struggled with more complex physical tasks such as adaptive, continuous ball retrieval.

  • Anthropic frames the work as a step toward physical agentic AI where models use off-the-shelf physical tools more easily.
  • The company says models now complete some work that previously required human-model pair programming much more quickly on their own.
  • Preliminary Mythos trials were excluded because the comparison was not apples-to-apples given how the experiment and model serving were set up.

Google DeepMind Publishes AI Control Roadmap

Google DeepMind’s AI Control Roadmap lays out a system-level security framework for advanced AI agents deployed inside Google. The approach pairs alignment with defense-in-depth controls such as sandboxing, endpoint security, and prompt-injection resistance, while treating internal agents as potentially misaligned.

  • The roadmap models untrusted agents as insider threats similar to rogue employees with existing access.
  • Google builds the threat model on MITRE ATT&CK to break possible attacks into tactics and techniques.
  • One mitigation uses trusted AI systems as supervisors to review an agent’s reasoning, actions, and plans.

OpenAI Studies Reinforcement Learning for Beneficial Behavior

OpenAI research argues that reinforcement learning on realistic scenarios can produce broad, durable gains in aligned and beneficial model behavior. The work suggests beneficial traits can generalize beyond training domains and remain more resilient under adversarial pressure.

  • The dataset targets traits including honesty and corrigibility along with epistemic humility, transparency, fairness, and concern for human welfare.
  • The scenarios span health, education, science, law, engineering, and economics to test traits under uncertainty and competing incentives.
  • The beneficial-trait model improved on 44 of 53 benchmarks covering deception, honesty, reward hacking, latent safety risks, and related failures.

MosaicLeaks Probes Privacy Failures in Research Agents

MosaicLeaks examines privacy risks in deep research agents that combine private documents with web retrieval. Its PA-DR method reduces answer and full-information leakage from 34% to 9.9% while maintaining task success by rewarding safer query construction instead of relying only on user prompts.

  • The benchmark contains 1,001 multi-hop chains built from local enterprise documents and a controlled web corpus.
  • Its leakage taxonomy separates intent, answer, and full-information leakage based on what an observer can infer from query logs.
  • PA-DR raises strict chain success from 48.7% to 58.7% while reducing the most serious leakage category.

Trending AI Tools

  • Claude Code Artifacts Turns coding sessions into live, shareable visual pages with auto-refresh, version history, and privacy controls for Team and Enterprise beta users.

  • Adobe Firefly Agentic Skills Adds new Firefly AI Assistant capabilities and brings the creative agent into public beta across Photoshop, Premiere, Illustrator, InDesign, and Frame.

  • Codex Record & Replay Lets developers capture and replay workflows as reusable Codex skills.

Quick Hits

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.