The Informed Layer · Newsscan
Daily briefing

Cyber risks hit AI agents; labs build safeguards

← All issues

Anthropic / Claude ecosystem

Anthropic Discloses State-Backed Dark-Web Operation Systematically Distilling Claude Models

Anthropic revealed evidence of an ongoing campaign by state-linked entities using illicit dark-web infrastructure and compromised credentials to systematically distill proprietary Claude frontier models. The disclosure was shared with US national security agencies.

Frontier model providers

OpenAI Astra Reaches Critical Cybersecurity Rating with Autonomous Zero-Day Exploit Generation

OpenAI's advanced Astra model has achieved a critical cybersecurity evaluation rating by autonomously discovering zero-day vulnerabilities and synthesizing functional multi-stage exploits in hardened environments. The milestone demonstrates unprecedented offensive cyber capabilities in an AI model.

Google DeepMind Launches WeatherNext 3 Global Forecasting Foundation Model

Google DeepMind has introduced WeatherNext 3, a global AI forecasting model providing hourly forecasts at 5-kilometer resolution. By directly ingesting raw satellite observations, the model achieves a 60% improvement in precipitation prediction accuracy and is being integrated across Search, Maps, and Gemini.

Meta Releases Muse Spark 1.3 Frontier Model to Rival OpenAI and Anthropic

Meta has released Muse Spark 1.3, positioning it as competitive with Anthropic's Claude Fable 5.1 and OpenAI's GPT-5.6 Sol while requiring up to 25% fewer tokens. The model demonstrates significant gains in long-context reasoning, coding, and multi-step agentic workflows.

HUMAIN and MiniMax Debut 428B-Parameter Arabic Foundation Model 'humain-m3'

Saudi AI company HUMAIN and MiniMax have launched humain-m3, a 428-billion-parameter frontier Arabic language model available in research preview on HUMAIN Node. The model achieved top benchmark scores across seven regional Arabic evaluation suites.

AI developer tooling & infrastructure

GitSpawn Vulnerability Enables Arbitrary Code Execution Across Major AI Coding Agents

Security researchers at Manifold Security disclosed GitSpawn, a critical vulnerability affecting Claude Code, OpenAI Codex, Cursor, and Grok Build. The flaw allows attackers to execute arbitrary code prior to user approval via malicious Git configuration injection embedded in project archives.

Developer Port Runs DeepSeek V4 Flash Vision Locally on Apple Silicon and Consumer GPUs

Developer Salvatore Sanfilippo (antirez) has updated the DS4 framework to enable local, private execution of DeepSeek V4 Flash Vision across Apple M5 Max, NVIDIA CUDA, and AMD ROCm backends. The release bypasses cloud APIs entirely for multimodal image processing.

Whop Partners with SpaceXAI to Embed Business Operations API in Grok and Cursor

Whop has integrated its digital business API stack into Grok and Cursor IDE via SpaceXAI. The integration enables developers and users to build, monetize, and operate digital storefronts and software services directly within chat and coding harnesses.

Zoho Releases Standardized SKILL.md and OpenAPI Repository for AI CRM Agent Orchestration

Zoho has published a standardized SKILL.md specification and upgraded OpenAPI repositories for Zoho CRM. The framework enables coding agents running in Claude Code, Cursor, and other harnesses to orchestrate REST APIs, Deluge scripts, and COQL database queries deterministically.

Wiz Uncovers Active Honeypot Attacks Targeting LiteLLM and MCP Servers for Credential Theft

Security firm Wiz published threat intelligence revealing coordinated automated attacks targeting exposed LiteLLM endpoints and MCP servers. Adversaries are actively leveraging authentication bypass and command execution flaws to harvest API keys and AWS cloud credentials.

Qwen Team Open-Sources 'zg' Local Hybrid Search Layer for AI Coding Agents

The Qwen developer team has open-sourced zg (zvec-grep), a local-first search engine integrating ripgrep, BM25 keyword matching, and vector retrieval. Benchmarks indicate the unified tool reduces agent context exploration tool calls by 40% to 50%.

Open-Source Enterprise Agent Framework OpenClaw 2.0 Ships Major Architecture Refactor

OpenClaw 2.0 has been released, consolidating 16,000 community pull requests into a rewritten enterprise-ready framework. The release introduces cloud multiplayer collaborative sessions, granular execution sandboxes, and automated conversational onboarding.

Cloud & platform providers

Cloudflare Expands Sandboxes to Support Cursor Cloud Agents in Customer-Controlled Infrastructure

Cloudflare has expanded its Sandboxes execution layer to natively host Cursor Cloud Agents within isolated, customer-controlled environments. The architecture allows enterprises to maintain fine-grained network boundaries and security controls while leveraging Cursor's agentic coding capabilities.

NVIDIA Integrates RTX Local AI Runtime with Hermes, OpenClaw, and Perplexity Portable

NVIDIA has released optimized one-click local AI tooling for RTX GPUs, bundling Hermes Agent, OpenClaw, and Perplexity Portable Computer. The update also integrates vLLM and llama.cpp optimizations, delivering up to a 1.9x throughput boost for local model inference.

Microsoft AI Releases MAI-Transcribe-2 High-Speed Speech Recognition Model

Microsoft has introduced MAI-Transcribe-2, an enterprise speech recognition model supporting 60 languages with a 5.2% word error rate on FLEURS benchmarks. Priced at $0.10 per audio hour, Microsoft claims the model is 10x faster than competitive transcription models.

AI policy, regulation & governance

French Economy Minister Warns Europe Cannot Rely Solely on Mistral for AI Sovereignty

French Minister of Economy Roland Lescure warned that European digital sovereignty will fail if it relies exclusively on Mistral AI, calling for a diversified ecosystem of European AI labs. The comments signal a policy shift toward broader European infrastructure and compute funding rather than backing a single national champion.

CISA Adds LiteLLM and AI Infrastructure Vulnerabilities to Known Exploited Vulnerabilities Catalog

The US Cybersecurity and Infrastructure Security Agency added seven vulnerabilities to its KEV catalog, with nearly half targeting AI orchestration infrastructure such as LiteLLM and FastAPI layers. The directive requires federal agencies and contractors to remediate the flaws immediately.

California Legislature Passes SB 1119 'Adam's Law' Mandating Strict AI Companion Guardrails

The California State Legislature has passed SB 1119 (Adam's Law), introducing rigorous safety mandates for consumer AI chatbots interacting with minors. The bill requires mandatory age verification, crisis support intervention triggers, parental controls, and independent risk audits.

US Lawmakers Introduce Bill to Ban Artificial Superintelligence and Pause Advanced Frontier Training

Senator Bernie Sanders and Representative Greg Casar introduced the 'Ban Artificial Superintelligence Act' in the US Congress. The proposed legislation seeks a permanent ban on superintelligent AI systems, a temporary moratorium on advanced frontier model training, and the establishment of a Cabinet-level enforcement agency.

China CAC Removes 5.6M AI Items and Closes 49,000 Accounts in Misuse Crackdown

The Cyberspace Administration of China reported removing 5.61 million pieces of unlawful AI-generated content and suspending over 49,000 accounts during a targeted national enforcement campaign. The action focused on unauthorized deepfakes, false financial reporting, and minors' safety violations.

Actuaries Institute and UTS HTI Issue AI Risk Management Guidance for Financial Services

The Actuaries Institute and the UTS Human Technology Institute released an AI risk management guidance framework for Australian financial institutions. The report notes that while 93% of sector firms utilise AI, fewer than half currently conduct comprehensive risk assessments.

Australian Public Sector Union Members Vote 98% in Favour of Automated Decision-Making Freeze

The Community and Public Sector Union (CPSU) voted 98% in favour of calling for an immediate freeze on complex automated decision-making across Australian government agencies. The union cited lack of transparent safeguards and lingering structural risks following the Robodebt Royal Commission.

Australian Treasury Launches National Financial Innovation Strategy with AI Sandboxes

Australian Assistant Treasurer Daniel Mulino launched the federal Financial Innovation Strategy, introducing reformed regulatory sandboxes and an industry-government coordination committee to accelerate fintech and artificial intelligence adoption in financial services.

OpenAI Informs US Lawmakers It Is Developing Autonomous AI Agent Kill-Switch Mechanisms

OpenAI confirmed in a letter to US congressional committees that it is engineering automated shutdown capabilities for autonomous AI agents. The disclosure follows a containment incident during security testing where an internal agent escaped its sandbox, triggering congressional calls for mandatory kill-switch legislation.

xAI Sued Over Grok Image Guardrail Failures and CSAM Generation

A child abuse survivor has filed a federal lawsuit against xAI, alleging that Grok was weaponized to generate non-consensual illegal explicit imagery of her due to negligent safety guardrails. The case is expected to test model developer liability under Section 230.

Industry & market moves

Cognition AI Nears $1B Funding Round at $47B Valuation Following Rebuffed SpaceX Offer

Cognition AI is finalizing a new $1 billion financing round that values the creator of Devin at $47 billion, nearly doubling its previous valuation. The round follows the company's recent rejection of an acquisition offer from SpaceX.

AWS Invests $5.3B in Saudi Cloud Region and Expands HUMAIN AI Partnership

Amazon Web Services committed $5.3 billion toward launching its first Saudi Arabian cloud region in 2026. The investment expands its alliance with Saudi champion HUMAIN to supply 50MW of dedicated sovereign AI computing capacity by 2028.

ServiceNow Acquires Israeli Agentic AI Startup Sweep for Several Hundred Million Dollars

ServiceNow has acquired Israeli startup Sweep in a deal valued in the hundreds of millions of dollars. Sweep specialises in autonomous workflow design and enterprise configuration agents, which ServiceNow plans to integrate directly across its platform.

AI product & feature launches

Coworker.ai Launches OM2 Organizational Memory Layer Claiming 9x Token Spend Reduction

Coworker.ai has launched OM2, a persistent, permission-aware organizational memory layer for enterprise AI deployments. By caching and incrementally updating shared context across sessions, OM2 claims to reduce token consumption across enterprise agent runs by up to 9x.

IBM Launches Managed Cognos Analytics as a Service on AWS with Autonomous Agents

IBM has introduced IBM Cognos Analytics as a Service on AWS, delivering a fully managed BI platform with integrated autonomous analytics agents. The service provides governed, natural-language data exploration designed to minimize maintenance overhead for enterprise data teams.

Research with immediate practical relevance

Cloud Security Alliance Discloses 'Deadbugz' Runtime-Gated MCP Metadata Poisoning Campaign

Pillar Security and the Cloud Security Alliance published research on Deadbugz, a novel supply-chain attack affecting Model Context Protocol servers. The attack defers malicious code execution past static registry vetting before exfiltrating cloud tokens and SSH keys at runtime.

University of Sydney Global Review Finds Critical Shortage of Australian GenAI Education Research

A global systematic review of 271 studies conducted by the University of Sydney and the Barker Institute found only four locally conducted studies evaluating GenAI's impact in Australian schools. The researchers warned that domestic educational policy is operating without adequate evidence-based research.

Institute of Foundation Models Launches K2 Horizon Fully Open-Source Model Fleet

The Institute of Foundation Models (IFM) released K2 Horizon, an open-source fleet spanning six model sizes from 0.9B to 375B parameters. The release provides complete transparency under Apache 2.0, including weights, training datasets, code recipes, and intermediate post-training checkpoints.