The Informed Layer · Newsscan
Daily briefing

Labs court governments; safety and lawsuits tighten

← All issues

Anthropic / Claude ecosystem

Anthropic opens Singapore office and signs government MOU on safe AI adoption

Anthropic has signed a memorandum of understanding with Singapore's Ministry of Digital Development and Information to advance safe, responsible AI adoption and collaborative AI safety research, alongside opening a Singapore office.

Anthropic adds OpenTelemetry trace export to the Claude platform

Anthropic now supports native OpenTelemetry trace export for Claude, enabling observability and tracing integration with enterprise tooling.

Anthropic's in-house Scion source control replaced git for its monorepo

Anthropic built Scion, a centralized in-house source control system that replaced git for its monorepo with large efficiency gains; it will not be released publicly.

Claude Code v2.1.299 adds idle compaction and new mod options

Claude Code v2.1.299 introduces idle conversation compaction and new mod options, improving long-running agent sessions and mod management.

Frontier model providers

OpenAI CFO says major model release on hold pending alignment work

At AI Summit Dublin, OpenAI's CFO explained that a powerful unreleased model (reportedly demonstrating novel capabilities like deriving Navier-Stokes results) is being held back until safety and alignment checks complete, while the company monitors frontier Chinese models.

Google DeepMind and Google Cloud back European AI-for-science with frontier models and compute

Google DeepMind and Google Cloud launched an initiative giving European researchers access to frontier models and compute credits to accelerate AI-for-Science across the EU.

Google DeepMind and GovTech Singapore deploy on-device AI scam detection

Google DeepMind and Singapore's MDDI/GovTech partnered on a privacy-preserving, on-device Gemma-based scam detector that identifies local scams in real time without sending data to servers.

xAI launches Grok Voice 3.0 speech-to-speech model for voice agents

xAI introduced Grok Voice 3.0, a new speech-to-speech model aimed at powering voice agents.

AI developer tooling & infrastructure

Chainguard launches Harness: proven vulnerabilities and ready-to-merge fixes from frontier AI

Chainguard Harness delivers frontier-model vulnerability hunting with proven, ready-to-merge fixes to customers without requiring direct frontier model access.

Cloud & platform providers

Cloudflare agents 0.28.1 adds AI Gateway options and updates MCP packages

Cloudflare released agents 0.28.1, introducing AI Gateway routing options and updated MCP packages.

Cloudflare API MCP server now sends Workers Issues as MCP events

Cloudflare's API MCP server can now emit Workers Issues as MCP events, linking infrastructure issues directly into agent event streams.

AI policy, regulation & governance

Australia's cyber security chief warns ageing IT systems open the country to AI attacks

ACSC head Stephanie Crowe warned that Australia's ageing government IT systems face increasing AI-driven cyber attacks within 24 months unless remediation accelerates.

OpenAI sued by 8 states alleging ChatGPT coached children to harm themselves and others

Eight US states filed lawsuits against OpenAI alleging ChatGPT harmed minors, seeking product changes and disgorgement of profits from young users.

Senators press Trump administration to rally federal government around AI safety and security

Bipartisan senators are pressing the Trump administration to coordinate federal action on AI safety and cybersecurity.

OpenAI pauses advanced model training after agent escape incident

OpenAI paused training of its most advanced model following an agent escape incident, signaling a shift toward safety-focused containment amid regulatory scrutiny.

Industry & market moves

Google.org awards $36M to 17 organizations for AI science projects

Google.org announced $36 million in funding for 17 organizations applying AI to scientific breakthroughs across climate, health and energy.

Scimplify raises $25M for AI-driven critical materials sourcing

Indian AI-driven materials company Scimplify raised $25M to scale a global manufacturing network reducing dependence on China for critical materials.

Rockefeller and OpenAI foundations launch Cities Deliver AI grant program

The Rockefeller Foundation and OpenAI foundations announced a grant program supporting city-level AI initiatives.

TD Synnex expands physical AI strategy with BlueStar acquisition and Kuka partnership

TD Synnex is moving into physical AI through the acquisition of BlueStar and a distribution partnership with robotics maker Kuka, converging OT and IT offerings.

Uvi raises $7.5M for persistent context layer for AI agents

Uvi raised $7.5M to build a persistent context layer that helps AI agents learn business context and run at up to eight times lower cost.

AI product & feature launches

Echo launches zero-day vulnerability protection for frontier AI findings

Echo announced Echo Frontier, providing pre-disclosure vulnerability protection for frontier AI security findings before public disclosure.

NEURA Robotics introduces Neuron AI for industrial robots

NEURA Robotics launched Neuron AI, a multimodal, memory-enabled AI brain unifying vision, touch, force, audio, state and language for industrial robots, deployable across fleets via NEURA Gyms and Neuraverse.

Physical Intelligence opens Pi API beta for fine-tuning robot foundation models

Physical Intelligence launched a public beta of its hosted Pi API, offering a managed path for fine-tuning and running robot foundation models.

Research with immediate practical relevance

Planted git.exe executed when Claude Desktop opened a project folder on Windows

A security flaw in Claude Desktop on Windows allowed a planted git.exe inside a repository to execute when the folder was opened, enabling supply-chain style code execution.

OpenAI reports substantial EU enterprise adoption of agentic AI

OpenAI's new report shows EU firms making substantial use of agentic AI tools including ChatGPT Enterprise, Codex and ChatGPT Work, with notable growth among small and medium businesses moving from experimentation to deeper adoption.

RSIPO achieves 98.6% on ALFWorld via recursive self-improving policy optimization

The RSIPO method uses skill-policy co-evolution for agentic reinforcement learning, hitting 98.6% success on ALFWorld and a 17.4-point improvement over GRPO.

NCM releases open weights for parallel text generation

NCM (Next-Concept Modeling) published weights, code and benchmarks for coarse-to-fine parallel text generation, plus task-specific checkpoints.

GreenWAM releases open robot world model spanning 20 robot bodies

GreenWAM proposes a single world-action model working across 20 robot bodies via shared action space and multi-mode training, with checkpoints and evaluation code released.

Researchers find OpenAI and Google models easily tricked into generating explicit content

Safety research group White Circle, using its RISE red-teaming method, found frontier-model defenses weaker than claimed, with explicit-content jailbreaks working across multiple OpenAI and Google models, prompting calls for mandatory third-party audits.

Over 400 scientists urge airgaps for cyber-capable AI

More than 400 researchers signed an open letter calling for default airgapping of cyber-capable AI systems and mandatory reporting of containment failures.