The Informed Layer · Newsscan
Daily briefing

Anthropic raises privacy bar; Agents gain ground; EU…

← All issues

Anthropic / Claude ecosystem

Claude Identity Verification Starts July 8: What Facial Data Anthropic Collects

Anthropic will implement biometric identity verification for Claude users starting July 8, becoming the first major US frontier AI lab to codify facial data collection (government ID, facial geometry scan, selfie) in its consumer privacy policy. This move raises significant data retention and security concerns regarding the third-party vendor Persona's surveillance capabilities.

Frontier model providers

Elon Musk Unveils Grok 3 as xAI Targets AI Supremacy; Perplexity Launches Deep Research Tool and Meta Ventures into Humanoid Robotics

xAI has launched Grok 3, claiming it to be the 'smartest AI on Earth,' following its training on 100,000 Nvidia H100 chips via the Colossus supercomputer. This release marks xAI's continued aggressive push for dominance in the frontier AI space.

MiniMax M2 & Agent: Ingenious in Simplicity - MiniMax News | MiniMax

MiniMax has launched its new M2 model, an open-source offering specifically optimized for AI agents and coding tasks. The company claims M2 operates at just 8% of Claude Sonnet's price while delivering double the inference speed, positioning it as a highly efficient and cost-effective solution.

AI developer tooling & infrastructure

每日一个开源项目（第136篇）：OpenMemory - 给 AI Agent 真正的认知记忆引擎-CSDN博客

OpenMemory, a new open-source cognitive memory system for AI agents, has been released. It utilizes a five-sector hierarchical decomposition model with temporal knowledge graphs, offering a distinct approach from traditional vector database RAG (Retrieval Augmented Generation) methods to provide more robust and persistent agent memory.

7,000 Langflow servers are under attack. LangGraph and LangChain have the same holes - CrypThinks

Active exploitation of remote code execution vulnerabilities across three major AI agent frameworks—Langflow, LangGraph, and LangChain—has been reported. This widespread attack highlights common AppSec blind spots in framework-layer security, where conventional WAF and EDR tools are unable to detect sophisticated attacks, posing a systemic risk to AI agent deployments.

Stop Making Your AI Coding Agent Grep Your Whole Repo — Try codebase-memory-mcp - DEV Community

A new open-source Model Context Protocol (MCP) server, codebase-memory-mcp, has been developed to significantly reduce token consumption for AI coding agents. It achieves this by indexing codebases into persistent knowledge graphs, eliminating the need for agents to repeatedly explore files during development, and reportedly cuts token usage by 99%.

Cloud & platform providers

No significant new developments.

AI policy, regulation & governance

EU Set to Brand AWS and Azure as Cloud Gatekeepers, Forcing Egress Fee Overhaul

The European Union is formally designating cloud infrastructure services, specifically AWS and Microsoft Azure, as digital gatekeepers under the Digital Markets Act (DMA) for the first time. This designation will trigger mandatory obligations on egress fees and API interoperability, aiming to foster greater competition and fairness in the cloud market.

Supreme Court Curtails Agency Power to Mandate AI Audits Without Clear Congressional Authorization — Credence Wire

The Supreme Court has struck down a federal agency mandate for algorithmic hiring audits, ruling that such mandates require explicit congressional authorization. This decision significantly reshapes the ability of federal agencies to regulate private-sector AI without clear legislative backing.

Here’s what you need to know about Canada’s new policies on social media and AI | 980 CJME

Canada has introduced three significant digital regulation bills: Bill C-34, which proposes a social media ban for individuals under 16; Bill C-36, a privacy bill establishing privacy as a fundamental right; and Bill C-16, which criminalizes non-consensual deepfakes. These bills represent a comprehensive effort to address emerging digital and AI-related harms.

House Appropriators Split on $12 Billion AI Research Package as Shutdown Clock Ticks — Credence Wire

US House and Senate appropriators are deadlocked over the allocation of a $12 billion AI research budget. A key point of contention is whether to divert half of the funds toward Pentagon applications, potentially delaying university safety grants and risking a government shutdown ahead of the September 30 deadline.

Six Weeks to EU AI Act Enforcement as Enterprises Stay Unprepared

With the EU AI Act's August 2, 2026, enforcement deadline for Annex III high-risk systems just six weeks away, a staggering 78% of affected organizations have yet to take meaningful compliance steps. Many lack basic AI system inventories, signaling widespread unpreparedness for the stringent requirements.

Chalmers refuses to release AI guardrail gap analysis

Australian Treasurer Jim Chalmers and the Albanese government have refused to publicly release the findings of a commissioned regulatory gap analysis on AI guardrails. The decision to withhold the report has drawn criticism regarding transparency in AI policy development.

Anthropic's Mythos AI Cracked 'Almost All' US Classified Systems in Hours, NSA Chief Reportedly Said | IBTimes UK

According to a reported account by the NSA director, a commercial AI model developed by Anthropic, named Mythos, allegedly penetrated nearly all US classified networks within hours during testing. This incident highlights the significant security implications and potential vulnerabilities posed by advanced AI systems.

Industry & market moves

Nobel Prize AI researcher leaves Google to Anthropic - Taipei Times

John Jumper, the 2024 Nobel Prize winner in chemistry and lead of the AlphaFold team, has departed Google DeepMind to join Anthropic. This high-profile move signals the escalating talent war within the competitive AI research landscape.

France Expands AI Infrastructure with NVIDIA, Including a 44 MW Data Center - Wedoany

France is actively operationalizing its AI infrastructure ambitions, with Mistral's 44 MW data center now live and the Campus AI expansion underway. European companies are deploying AI agents in production across various sectors, including healthcare, telecom, manufacturing, and energy, following commitments made at GTC Paris a year prior.

Dream Raises $260 Million At $3 Billion Valuation For Sovereign AI Platform

Dream, a sovereign AI platform specializing in secure AI for governments, has raised $260 million in Series funding at a $3 billion valuation. The funding will be used to expand its AI and cyber defense capabilities globally, focusing on national security applications.

Jack Technology and Siemens partner to bring AI and humanoid robots into apparel manufacturing

Jack Technology and Siemens have partnered to deploy AI orchestration software and humanoid robots within apparel manufacturing facilities. This collaboration targets a 30% efficiency gain across production and development workflows, leveraging Siemens Intelligence Center X, Mendix low-code platform, and Designcenter engineering software integrated with Jack Technology's robots.

OpenAI Secures 4.5 Gigawatts with Oracle as Meta and SSI Battle for Superintelligence Dominance - AI Tech News Today

OpenAI has finalized a $30 billion cloud computing deal with Oracle, securing 4.5 gigawatts of computing capacity to power its ambitious Stargate AI infrastructure. This massive investment underscores the escalating demand for energy and compute resources in the race for AI supremacy.

AI product & feature launches

Meta expands AI-powered teen safety protections across its platforms | Digital Watch Observatory

Meta is globally expanding its AI-based age detection and default age-appropriate content protections for teen users. This initiative shifts towards automated safeguards, including a 13+ default content setting, parental notifications, and Family Center integration, reducing reliance on manual parental configuration.

Cyberattacks on Civil Society Surge Worldwide as Journalists, Human Rights Groups Face Growing Digital Threats: Cloudflare Report - Digital India Times

Cloudflare's Project Galileo 2026 Report reveals a 44% surge in cyberattacks targeting civil society organizations globally, with media outlets facing disproportionate targeting. The report also highlights an increase in government-linked internet shutdowns used to suppress dissent.

Alibaba ATH Launches HappyOyster, a Real-Time Interactive World Model - Wedoany

Alibaba ATH has launched HappyOyster 1.0, a new real-time interactive world model that enables users to explore, interact with, and direct dynamically evolving digital environments. This model addresses the limitations of static text-to-video generation by allowing for fluid, user-driven narrative and environmental changes.

China Galaxy General-Purpose Robot Releases Humanoid Robot Motion Control Model Based on 20,000 Hours of Data - Wedoany

Galaxy General-Purpose Robot in China has unveiled AstraBrain-WBC 0.5, a humanoid robot motion control model that applies a GPT-like large-scale training paradigm. Trained on the industry's largest humanoid robot dataset (20,000 hours), it achieves general-purpose motion foundation model capabilities for real-time control.

Research with immediate practical relevance

A Microsoft researcher built a goat-powered LLM in Age of Empires II to prove it's not sentient

Microsoft researcher Adrian de Wynter has demonstrated that the core functionalities of a Large Language Model (LLM) can be replicated using video-game mechanics, specifically by employing 'goats as NAND gates' within Age of Empires II. This research argues that the human-like tone of LLMs reflects their presentation rather than genuine sentience.

Humans actively undermine AI lie detectors because they don't want to accuse people of lying

A new study by Riccardo Loconte et al. from the IMT School for Advanced Studies Lucca reveals that humans actively undermine high-accuracy AI lie detectors by consistently rejecting confident deception accusations. This human tendency to avoid accusing others of lying degrades the AI system's performance by 14 percentage points.