The Informed Layer · Newsscan
Daily briefing

AI advances, agents scale; governance tightens

← All issues

Anthropic / Claude ecosystem

Anthropic accuses Alibaba of largest known model distillation campaign using 25,000 fraudulent accounts

Anthropic disclosed in a congressional letter that operators affiliated with Alibaba conducted a large-scale model distillation attack generating 28.8 million exchanges with Claude through nearly 25,000 fraudulent accounts between April 22 and June 5, 2026. The campaign targeted Claude's agentic reasoning, software engineering, and long-horizon task capabilities, escalating geopolitical AI IP theft and triggering calls for stronger export controls.

US Commerce Department suspends Anthropic's Fable 5 and Mythos 5 globally via export control

On June 12, 2026, the US Department of Commerce issued an export control directive ordering Anthropic to suspend access to Fable 5 and Mythos 5 for all foreign nationals, citing a jailbreak vulnerability as a national security risk. Unable to filter by nationality at API scale, Anthropic disabled both models globally within hours, establishing the first export control enforcement action against a deployed commercial AI model.

Frontier model providers

OpenAI unveils Jalapeño, custom AI inference processor built with Broadcom

OpenAI announced Jalapeño, a custom AI chip designed specifically for LLM inference, built in collaboration with Broadcom in nine months. The chip shows superior performance-per-watt versus existing top chips and is already running GPT models in the lab, representing OpenAI's strategic pivot toward vertical integration of hardware and reducing API infrastructure dependency.

Z.ai releases GLM-5.2 open-source frontier model at one-sixth the inference cost of US models

Chinese AI startup Z.ai released GLM-5.2, a 744-billion-parameter open-weight model with frontier-class performance on reasoning and coding benchmarks, while operating at roughly one-sixth the inference cost of leading U.S. frontier models. The release tilts the competitive landscape during heightened geopolitical AI tension and signals narrowing capability gaps between open and closed models.

Meta remains only major U.S. AI lab refusing federal frontier model security review

Meta is the sole major U.S. AI developer that has not voluntarily agreed to submit its frontier models to the Trump administration's federal safety review program. OpenAI, Anthropic, Google, xAI, and Microsoft have all formalized agreements with the Center for AI Standards and Innovation; Meta's isolation amid heightened government scrutiny creates strategic risk.

Meta hires three Virtue AI founders including Dawn Song to lead AI safety at Superintelligence Labs

Meta Superintelligence Labs announced the hiring of three Virtue AI co-founders—Bo Li (CEO), Dawn Song (co-founder), and Sanmi Koyejo—to lead AI safety and security work on frontier models. The acquisition of independent red-teaming expertise inside a frontier lab concentrates evaluation capability while accelerating safety integration into model development.

Meta deploys Muse Spark frontier model to content moderation, replacing majority of human reviewers

Meta accelerated its transition to AI-powered content moderation, with large language models now handling approximately 50% of content review requests, with targets exceeding 90% for specific categories by year-end. The company recently switched from Google's Gemini to its own Muse Spark foundation model, demonstrating large-scale agentic deployment to offset AI infrastructure costs.

Cloud & platform providers

AWS Bedrock achieves FedRAMP High and DoD IL-4/5 certification for government AI workloads

Amazon Bedrock, AWS's managed foundation model service, completed FedRAMP High Authorization and DoD IL-4/5 certification, enabling government agencies to deploy managed AI models in secure environments. The compliance milestone signals AWS's commitment to the government AI market amid expanded federal oversight.

Qualcomm and Hugging Face expand partnership for edge-to-cloud AI deployment on Dragonfly

Qualcomm and Hugging Face announced an expansion of their strategic relationship enabling the Hugging Face community of 16 million developers to deploy models on Qualcomm's Dragonfly data center solutions and Snapdragon devices, with agentic AI orchestration across hybrid edge-to-cloud environments.

AI policy, regulation & governance

India's RBI mandates kill-switches and human oversight for all AI in banking

The Reserve Bank of India released draft guidance on June 25 establishing mandatory Model Risk Management frameworks for AI and machine learning in banking. The rules require board-level oversight, independent validation of third-party models, mandatory kill-switch provisions, and explicit human override mechanisms, raising governance expectations across Asian financial AI deployments.

Australia signals AI reform agenda embedding 'national values of fairness' amid Five Eyes cyber warning

On June 25, Prime Minister Anthony Albanese signaled major AI reform as next priority, stating the government will embed 'national values of fairness' in AI rollout while protecting sovereignty. The announcement follows a Five Eyes intelligence alliance warning that frontier AI models pose faster, more complex cyber threats within months, not years.

South Korea launches inter-agency task force to combat AI-enabled crime across 10 government agencies

South Korea formally established the Inter-Agency Consultative Body for Responding to AI Crime on June 26, comprising 10 government agencies to address crimes abusing AI including deepfake sexual exploitation, false advertising, and financial fraud. The multi-ministry coordination model establishes enforcement precedent across financial, platform, telecommunications, and law enforcement sectors.

Canada tables Bill C-36 establishing digital regulator with AI transparency mandates and enforcement powers

Canada's Minister of AI tabled Bill C-36 on June 15, proposing the Protecting Privacy and Consumer Data Act to establish the Digital Safety and Data Protection Commission with enforcement powers up to CAD $10 million or 3% of global revenue. The bill explicitly requires transparency disclosures for automated decision-making systems and AI-driven decisions.

Industry & market moves

onsemi acquires Synaptics for ~$7B to emerge as Physical AI infrastructure provider

onsemi announced a definitive agreement to acquire Synaptics in an all-stock transaction valued at approximately $7 billion. The combination extends onsemi's power and sensing capabilities to intelligent systems at the intersection of Physical AI, positioning the company for autonomous vehicles, robotics, and AR/VR applications during accelerated embodied AI scaling.

Adobe acquires Topaz Labs for AI image and video enhancement

Adobe announced an agreement to acquire Topaz Labs, the Emmy-winning maker of AI image and video enhancement tools. The deal brings upscaling, restoration, and on-device AI capabilities into Creative Cloud, Photoshop, Lightroom, and Premiere products, leveraging Topaz's Neurostream technology for local execution on consumer hardware.

General Intuition closes $320M Series B at $2.3B valuation for embodied AI scaling

General Intuition announced the closing of its $320M Series B round at a $2.3B valuation, bringing total disclosed capital to $454M. The round is led by Khosla Ventures and includes investors Jeff Bezos and Eric Schmidt, primarily funding compute scaling and API expansion for the company's world models trained on video game action data for robotics and embodied AI.

EXL acquires iMerit for $310M to strengthen AI training and model evaluation

ExlService announced a definitive agreement to acquire iMerit for up to $310M ($170M upfront plus $140M in earnouts). iMerit provides AI model training, evaluation, and reinforcement learning services across multiple modalities, strengthening EXL's enterprise AI capabilities and access to foundation model builder relationships.

AI product & feature launches

Naver officially launched AI Tab on June 26, expanding conversational AI search capabilities across its platform and replacing the 'Green Dot' search interface that had been a fixture since 2018. The launch demonstrates integrated AI search with shopping, reservations, and place discovery, with search share rising from 63.82% to 66.34% after beta.

Fortinet launches FortiSOC, unified cloud-native SOC platform with agentic AI

Fortinet launched FortiSOC, a cloud-delivered security operations center platform unifying SIEM, SOAR, threat intelligence, ITDR, UEBA, and agentic AI into a single software-as-a-service experience. The platform consolidates six core security operations functions, reducing tool sprawl and enabling AI-driven investigation acceleration.

Salesforce introduces Help Agent with outcome-based monetization for customer service

Salesforce announced Help Agent, a prebuilt service agent set atop the Agentforce platform, available July 2026. The agent can be connected to company knowledge, actions, and communication channels in minutes and introduces outcome-based monetization ($2 per resolved customer issue) rather than token-based pricing, signaling industry shift toward efficiency-based models.

Dragos launches EmberAI, OT-native AI assistant for critical infrastructure security

Dragos launched EmberAI on June 23, an AI assistant purpose-built for operational technology security teams defending power grids, factories, water systems, and pipelines. The system correlates security alerts against operational context and known OT threat behavior, bringing vertical domain expertise into agentic workflows.

Research with immediate practical relevance

Qwen releases AgentWorld, language-based world model for synthetic agent training

Qwen published Qwen-AgentWorld, a language world model capable of predicting terminal outputs, browser states, phone screens, and other text-representable environments. The system enables agents to train against environment predictions without live infrastructure, achieving skill transfer from invented worlds to real tasks (e.g., 34→50 score on real web search after synthetic training).

Unconventional AI releases Un-0, image generation via coupled oscillators substrate

Unconventional AI released Un-0, an open-weights image generation model built on simulated coupled Kuramoto oscillators rather than conventional neural networks. The architecture is designed for eventual physical oscillator-based hardware deployment with claimed 1,000× energy efficiency gains, demonstrating AI workloads can map to fundamentally different computing substrates.

Inclusion AI open-sources Ling and Ring 2.6 trillion-parameter agentic models

Inclusion AI released open-source Ling-2.6 and Ring-2.6 models balancing instant response generation with deep reasoning via dual-model architecture. The system addresses the speed-vs-complexity tradeoff in agentic LLMs with efficient attention mechanisms, enabling practical deployment on consumer hardware (160 tokens/second on 8GB VRAM).

DeepReinforce releases Ornith-1.0, coding models that learn their own RL scaffolds

DeepReinforce released Ornith-1.0 (9B to 397B variants, MIT-licensed) that learn to generate their own agent harnesses during RL training rather than relying on hand-engineered orchestration. The 397B model scores 82.4 on SWE-Bench Verified (second only to Claude Opus at 87.6) and demonstrates agentic scaffolding can be learned rather than designed.