The Informed Layer · Newsscan
Daily briefing

Anthropic doubles down on safety; AI rules advance

← All issues

Anthropic / Claude ecosystem

Anthropic publishes 2026 Usage Policy update covering high-risk Claude use cases

Anthropic released its 2026 usage policy update, clarifying and expanding rules around Claude's newer capabilities and high-risk use cases, including explicit prohibitions on model abuse and election interference. The update aligns policy with enforcement against misuse of agentic and high-stakes features.

Anthropic expands support for US scientific discovery via Genesis Mission

Anthropic announced it is expanding its commitment to American scientific research by funding AI-enabled discovery through the Genesis Mission across multiple US agencies.

Anthropic launches free opt-in vulnerability-finding service for open-source software

Anthropic launched OSS Scanner, a free opt-in service that lets open-source projects use Claude to rapidly discover vulnerabilities, generating fully model-generated reports for maintainers.

Anthropic launches AI push to protect critical infrastructure

Anthropic is launching a Critical Infrastructure Defense Program to deploy its top AI models and engineers to bolster cyber defences for power grids, water utilities and other critical infrastructure.

Claude Code 2.1.293 reverts a two-day-old cloud-session fix

Anthropic shipped Claude Code v2.1.293, reverting a cloud-session fix introduced in v2.1.290 two days earlier, amid a rapid sequence of fixes and reversions highlighting unstable cloud-session behaviour across recent releases.

Frontier model providers

No significant new developments.

AI developer tooling & infrastructure

JetBrains launches Mellum 2.1, a fast open model for coding agents

JetBrains released Mellum 2.1, a fast open model for coding agents trained with reinforcement-learning-in-the-loop post-training and offering private self-hosted deployment options.

Microsoft launches local on-device MAI Code 1.1 Flash coding model

Microsoft brought local models and sandboxed tools to Windows and GitHub Copilot with MAI Code 1.1 Flash, a smaller-footprint model using quantization and speculative decoding for responsive coding agents on Windows hardware.

Docker launches YAML-defined AI agent orchestration tool

Docker introduced a declarative AI agent orchestration tool that lets teams define agents in YAML and run them on container infrastructure without writing code.

Atlassian MCP expands AI agent reach across more products

Atlassian expanded its MCP server to cover more Atlassian products and workflows, improving AI agent execution with leaner context and up to 25% token savings in benchmarks.

Cloud & platform providers

Grok 4.7 goes GA in Microsoft Foundry with OpenAI-compatible APIs

Grok 4.7 reached general availability in Microsoft Foundry on Azure, callable via OpenAI-compatible APIs so Azure customers can deploy the model with familiar OpenAI clients, with guidance to evaluate behaviour per workload.

Carbon-A gene-annotation model predicts coding regions across diverse genomes

Hugging Face released Carbon-A, a model claimed to be the first to predict protein-coding regions directly from DNA with a single model across diverse life forms, using a 98,304-base-pair context window and outperforming several baselines.

AI policy, regulation & governance

India to release draft AI regulations within a month

India's IT Minister Ashwini Vaishnaw said the central government will publish draft AI regulations and solicit public comments within a month, signalling a forthcoming national AI regulatory framework covering deepfakes and IT governance.

Colorado releases revised draft rules for ADMT Act and Conversational AI Act

The Colorado Department of Law released the newest revised draft rules for its Automated Decision-Making Technology (ADMT) Act and Conversational AI Service Rules on October 6, advancing the state's rulemaking process.

Australia moves toward bank-style risk-management duties for frontier AI labs

Assistant Technology Minister Andrew Charlton said AI firms will carry the safety burden under new Australian laws, pursuing a 'systems regulation' model that requires AI companies to run risk-management processes and prove safety systems work, drawing on banking and aviation oversight models rather than prescribing hard-coded rules. Legislation is likely early next year.

FTC advances AI safety probe toward civil investigative demands

The FTC appears poised to issue civil investigative demands to frontier AI firms, marking a formal moving-enforcement phase of its AI safety probe ahead of any lawsuit decision.

Industry & market moves

Meta donates 1,000 AI glasses to Singapore's disability community

Meta is donating 1,000 Ray-Ban Meta AI glasses to Singapore's disability community, coupled with a training grant, to improve independence for users with disabilities.

Manus raises over $500M in first round since Meta split

Chinese AI startup Manus (Butterfly Effect) raised over $500 million in its first funding round since a regulatory unwind of a $20 billion Meta deal, signalling continued strong capital inflows and a potential path toward IPO.

GreyOrange acquires Site Bionics to accelerate physical AI strategy

Warehouse robotics firm GreyOrange acquired Site Bionics to leverage its digital twin platform and RFID-enabled insights, expanding its physical AI strategy for warehouse operations.

Nous Research raises $90M at $1.5 billion valuation

Nous Research closed a $90 million funding round at a $1.5 billion valuation to scale its open-source AI foundation models and infrastructure.

AI product & feature launches

Windows brings DeepSeek V4 Flash, Nemotron and llama.cpp to local AI on Windows

Microsoft announced DeepSeek V4 Flash and Nemotron models running locally on Windows, along with llama.cpp support coming to Windows ML, expanding on-device open-weight AI tooling.

Research with immediate practical relevance

Australia's AI Safety Institute releases report on safe oversight of advanced AI systems

Australia's AI Safety Institute published a new report on epistemic safety in scalable oversight, introducing frameworks and benchmarks to study scalable AI oversight, informing the National AI Plan and international alignment research.

KAIST SQAM replaces per-step Jacobians in flow-policy RL with closed-form adjoint

KAIST researchers introduced Scalar Q-Adjoint Matching (SQAM), which replaces per-step Jacobians in flow-policy reinforcement learning with a closed-form scalar adjoint, achieving 75% on OGBench with higher benchmark and real-robot gains.

OpenAI says Russian influence operation targeting Latin America used ChatGPT

OpenAI publicly attributed a broad Russian influence campaign spanning multiple countries to misuse of ChatGPT, marking a rare high-profile disruption of foreign influence activity using its platform.

OpenAI publishes October 2026 safety update for GPT-6 Sol and GPT-6 Luna

OpenAI's Deployment Safety Hub published its October 2026 model safety training and evaluation update covering GPT-6 Sol and GPT-6 Luna, detailing safety training and evaluation results for the current models.

AI agents design an openTPU chip that runs their own inference

Researchers demonstrated an AI-designed openTPU hardware accelerator that runs its own inference, validated for bit-exact behaviour across ISA simulator, RTL, and physical hardware, demonstrating end-to-end correctness for several models.