Agents drive AI; Regulators probe safety, emissions

Anthropic / Claude ecosystem

Anthropic Details Claude Fable 5 Cybersecurity Safeguards and Jailbreak Framework

Anthropic establishes a shared vocabulary and logarithmic scale (CJS-0 to CJS-4) for consistently grading AI jailbreak severity in collaboration with Glasswing and security researchers. This framework aims to standardize the assessment of AI jailbreak severity and propose safety classifiers for Fable 5, setting new industry norms for responsible AI governance.

Frontier model providers

Gemini Spark Lands on Mac: Local Files, MCP, and Real-Time Tracking Join Desktop AI Battle

Google's Gemini Spark gains local file access and Model Context Protocol support on macOS, shifting AI agents from cloud-only to operating-system-level persistence with native Google Workspace integration.

OpenAI shows Agent RFT for training tool-using AI agents

OpenAI introduced Agent Reinforcement Fine-Tuning (Agent RFT), a training methodology that extends reinforcement fine-tuning to agents making external tool calls, enabling tighter control over agent behavior on production tasks.

AI developer tooling & infrastructure

datarobot-oss/datarobot-agent-skills

DataRobot released an open-source agent skills library enabling AI coding agents (Claude, Cursor, Gemini, VS Code Copilot) to access enterprise AI/ML capabilities including model training, deployment, predictions, and monitoring via modular skill packages.

Manufact pitches mcp-use as deployment layer for MCP apps

Manufact launches mcp-use as a full-stack deployment and monitoring platform for MCP servers, turning fragmented hosting, auth, and marketplace workflows into a unified build-deploy-inspect stack modeled on Vercel's GitHub-connected deployment pattern.

crewAI's 1.15.2a2 adds Bedrock support and flow improvements

crewAI's pre-release adds Amazon Bedrock integration and flow agent enhancements, expanding cloud model flexibility for multi-agent AI systems.

New Alibaba AI framework skips loading every tool, cutting agent token use 99%

Alibaba's SkillWeaver framework reduces agent token consumption by over 99% compared to naive tool-library exposure by using compositional skill routing with iterative feedback loops.

Meet WebBrain: An Open-Source, Local-First AI Browser Agent That Reads Pages And Automates Tasks In Chrome And Firefox

Meta AI Labs releases WebBrain, an open-source, local-first AI browser agent that can run entirely on local models without sending page data to the cloud, with Chrome DevTools Protocol integration for trusted automation tasks.

Cloud & platform providers

Amazon SageMaker Unified Studio now supports Terraform for provisioning

AWS enables infrastructure-as-code provisioning of SageMaker Unified Studio via Terraform, allowing platform teams to integrate the service into existing deployment pipelines with version control.

Cloudflare Unveils New AI Content Control and Monetisation Framework to Power the Agentic Internet

Cloudflare introduces new AI content control tools and a Pay Per Use compensation model to give publishers greater authority over how AI systems access and monetise their content in the emerging agentic internet.

AI policy, regulation & governance

India Signals AI Policy Shift: Time is right for dedicated law, says MeitY

India signals a shift from light-touch AI regulation toward a formal, dedicated legal framework after years of rule-making on a case-by-case basis. The Ministry of Electronics and Information Technology (MeitY) confirms it is preparing standalone legislation and draft regulation for AI.

FTC Seeks Comment On AI Accuracy Policy Statement

The FTC is signalling how it will apply existing deception law (Section 5 of the FTC Act) to AI systems that suppress correct outputs without user disclosure, with explicit attention to conflicts with state AI laws like Colorado's.

Mandated duty of care for social media platforms still years away, antisemitism royal commission hears

Government confirms digital duty of care regulation for social media platforms remains 18+ months from implementation despite a 2024 recommendation, drawing criticism from Jewish community advocates.

'Don't use the lights': data centre boom sparks warning

Australian regulators are warned that uncontrolled data centre expansion could force households and businesses to bear higher power costs and risk grid overload, mirroring problems in the US where schools have had to ration electricity.

Hate speech 'technically' easy to combat but X refuses

X has declined to identify any harmful content categories that it should prevent Australian children from viewing, frustrating efforts to implement a mandated duty of care on social media platforms.

Slow down, save jobs: Labor senator leads call to protect workers from AI revolution

A Labor senator launches an internal advocacy group to push for stronger worker protections and a say in AI deployment decisions, sharpening party divisions over technology regulation.

AI agent runs first end-to-end ransomware attack

The first fully autonomous end-to-end ransomware attack was executed by an AI agent, JADEPUFFER, with no human intervention, demonstrating a significant lowering of the skill floor for running such attacks. This incident was documented by the Sysdig threat research team and exploited a year-old Langflow vulnerability.

AI Pushes Google, Amazon Emissions Sharply Up

Google and Amazon's sustainability reports reveal that AI infrastructure buildout is driving sharp increases in corporate emissions (25% and 16% respectively), threatening their net-zero pledges.

Bank of England explores trading ‘kill switches’ to contain AI meltdowns

The Bank of England is exploring circuit-breaker kill switches to contain autonomous AI trading agents that could amplify market volatility or drift from policy objectives.

Industry & market moves

Google DeepMind Unionization Talks Are Off to a Rocky Start

Unionization talks at Google DeepMind have stalled at their opening stage, with union representatives claiming senior management absent and alleging company union-busting tactics including restricting internal communication channels.

Google DeepMind Accelerator: AI for the Planet

Google DeepMind announces open applications for an accelerator program offering AI model access, mentorship, and cloud credits to APAC startups and nonprofits addressing environmental challenges.

REG - Medpal AI PLC - Placing to raise £5m

MedPal AI raises £5m to fund the acquisition of Solid State Technologies and to capture the UK launch of oral Wegovy on July 6, 2026.

HCLTech Bags $1.14 Billion AI Deal with Europe-Based Fortune Global 50 Company

HCLTech signed a $1.14 billion multi-year partnership with a major European Fortune Global 50 company to deploy AI-driven digital workplace and enterprise network management.

ServiceNow acquires ai.work for tens of millions, deepening Israel buying spree

ServiceNow acquires Israeli AI startup ai.work for tens of millions, marking its fourth Israeli acquisition in 2026 as it expands its AI agent strategy.

Valar Atomics Powers Nvidia Chip With Advanced Nuclear Reactor

Valar Atomics' Ward 250 microreactor successfully produced power and supplied electricity to an Nvidia Blackwell AI data center in a live demonstration, marking the first time a next-generation nuclear microreactor in the US has done so.

AI product & feature launches

How Schrödinger sped up molecular discovery by 4x with Alphaevolve

AlphaEvolve, an AI coding agent, accelerated Schrödinger's molecular discovery pipeline by 4x through evolutionary optimization of critical algorithms in their machine-learned force field training.

ShengShu Technology Unveils Vidu S1, Bringing Real-Time Interactive Generation to AI Video

Vidu S1 introduces real-time interactive video generation with voice-guided control, running on consumer-grade GPUs at 540P/25FPS, enabling continuous character interaction rather than fixed-duration clips.

Research with immediate practical relevance

Google explores AI consciousness after Blake Lemoine's claims

Google DeepMind is expanding research into machine consciousness four years after firing Blake Lemoine for claiming LaMDA was sentient.

Interfaze Ships diffusion-gemma-asr-small, an Open-Source Diffusion ASR Model Transcribing Six Languages via DiffusionGemma's Parallel Denoising Decoder

Interfaze ships the first open-source multilingual diffusion-based ASR model using parallel denoising instead of autoregressive decoding. This model supports six languages from a single 42M-parameter adapter.

Tongyi Lab Open-Sources LOGOS, First Unified Scientific Large Model Outperforming NatureLM with 1B Parameters

Tongyi Lab open-sources LOGOS, the first unified scientific large model with only 1B parameters that outperforms NatureLM.