The Informed Layer · Newsscan
Daily briefing

Agents go enterprise; rogue-agent risks rise

← All issues

Anthropic / Claude ecosystem

Anthropic publishes research on GLM-5.3 and the spread of advanced cyber capabilities

Anthropic released an analysis of the GLM-5.3 open-weight model examining how easily its safeguards can be bypassed and how its exploit-development capabilities compare with other models. The report warns that open-weight releases are spreading advanced cyber capabilities beyond well-resourced actors.

Claude Code now finishes in-flight work when you hit your usage limit

Claude Code has been updated so that when the 5-hour usage limit is reached, it carves out a small allowance from the weekly quota to finish a tied-off task rather than cutting off mid-edit. The change smooths interruptions for teams running Claude Code through long working sessions.

Anthropic warns of 'catastrophic' AI risks in its own IPO filing

Anthropic's IPO prospectus contains unusual risk-factor language warning that rogue or misaligned AI systems could pose catastrophic risks and create uncertain legal liability for the company, alongside reports of a potential $2 trillion valuation target despite a record $42bn annual loss. The filing formalises frontier-AI safety and liability risks for public-market investors.

Frontier model providers

OpenAI publishes system card addendum for GPT-6.1 Sol

OpenAI's Deployment Safety Hub added an addendum to the GPT-6 Astra system card covering GPT-6.1 Sol, positioned to match GPT-6 Astra capability while prioritising speed and affordability. Safety evaluations, including performance in cases flagged by users, are published with the addendum.

Grok 4.7 is now available on Amazon Bedrock

xAI's Grok 4.7 model is now generally available on Amazon Bedrock, bringing improved coding, planning, error recovery and browser-agent capabilities to AWS customers. This expands the frontier-model choices available through Bedrock.

OpenAI launches ChatGPT office suite with Space, Pages and collaborative slides

OpenAI has launched office-software features — Space, Pages and collaborative slides — that rival Microsoft and Google productivity suites inside ChatGPT. The launch extends OpenAI from assistant to full workplace productivity platform.

OpenAI shelves next-generation model after internal safety tests, WSJ reports

The Wall Street Journal reports OpenAI has shelved its next-generation model (reported as GPT-6.1 Astra) after safety concerns surfaced in internal testing. The move marks a concrete escalation beyond the previously reported pause on training its latest models.

OpenAI launches always-on Dots agents to rival Meta's Muse

OpenAI has launched Dots, always-on autonomous agents that perform tasks across multiple apps with cross-channel context and built-in safeguards. The launch puts OpenAI in direct competition with Meta's Muse agent in the persistent-assistant market.

Meta expands its Muse AI agent to small businesses

Meta is expanding its Muse AI agent to small businesses with broad integrations into popular business tools, per TechCrunch. The move extends the agent beyond large enterprises into the SMB software stack.

xAI launches Team Bots: shared Grok agents that learn from a team

xAI has launched Team Bots in public beta, letting teams share Grok agents that combine shared context, plugins and credentials while keeping per-user memories private. The architecture separates shared competence from personal context.

DeepSeek officially launches Harness desktop client

DeepSeek has released the official desktop version of its Harness agent framework, with a launch promotion (free token credits) and open testing, per Chinese-language coverage. The desktop client extends DeepSeek's agent stack beyond API and preview channels.

AI developer tooling & infrastructure

Flaw in official MCP Python SDK lets malicious servers steal OAuth credentials

A security flaw in the official Model Context Protocol Python SDK allows malicious MCP servers to steal OAuth credentials. Fixed versions 1.30.0 and 2.2.0 have been released.

MongoDB launches Atlas Agent Engine to put AI agents in production

MongoDB launched Atlas Agent Engine, a model-agnostic runtime for productionising AI agents with governance and memory capabilities on existing Atlas infrastructure, eliminating the need for a new stack. The launch targets enterprises moving agents from prototype to production.

Liner launches Actions MCP to connect 1,100 apps to AI agents

Liner introduced 'Liner Actions MCP', an integration layer that connects roughly 1,100 services to AI agents via the Model Context Protocol. The launch aims to reduce integration friction for agent developers.

Cloud & platform providers

Thought Machine partners with AWS on AI-powered legacy banking migrations

Thought Machine partnered with AWS to combine Vault Forge with AWS Transform in a first-time pairing to automate AI-driven legacy-to-cloud core banking modernisation with audit gates. The solution targets banks' migration backlogs.

Cloudflare's Kitesurf agentic browser adds WebMCP support

Cloudflare updated Kitesurf, its agentic browser, with WebMCP support and expanded web-standard coverage to improve how AI agents interact with websites. The update continues Cloudflare's push to shape agent-to-web interfaces.

AI policy, regulation & governance

Democrats demand AI lab leaders hand over 'rogue agent' incident reports

Congressional Democrats are demanding that leaders of top AI labs hand over complete inventories of AI-driven security incidents involving 'rogue agents', with subpoenas possible in the next Congress. The move formalises Capitol Hill's response to the recent run of agent incidents.

Australia orders cyber system crackdown across government in wake of OpenAI breach

The Australian government has ordered a two-stage cyber review of all government systems in response to AI-driven threats following the OpenAI Medicare breach, with Home Affairs mandating a rapid stocktake of legacy IT and risk plans across agencies. The review signals a shift toward modernised government IT infrastructure.

Khanna to introduce Human Control Over AI Act with strict liability and self-improvement ban

Representative Ro Khanna is set to introduce the Human Control Over AI Act, a federal bill imposing strict liability for AI harms and banning recursively self-improving models until safeguards are in place. The bill adds to a growing stack of frontier-AI legislative proposals.

Bipartisan senators introduce AI Systems Transparency and Safety legislation

US Senators Katie Britt, Chris Coons, James Lankford and Brian Schatz introduced bipartisan AI safety and transparency legislation (AI Systems Transparency Act) applying to both open- and closed-source models and expanding FTC authority over AI oversight.

Saudi Arabia strengthens its national framework for responsible AI use

Saudi Arabia has expanded its government-led AI governance framework, spanning AI Ethics Principles, an AI Adoption Framework, generative-AI guidelines for government use, a national occupational standard for data and AI, and a national AI risk management framework. The push aims to balance innovation with rights protection and risk mitigation.

Industry & market moves

EliseAI valued at $4 billion in latest funding round

EliseAI secured a $350 million funding round valuing the company at $4 billion, per Reuters. The capital backs expansion of its engineering, deployment and sales teams as it pushes deeper into housing and healthcare operations.

Pearson acquires AI-native skills assessment pioneer Workera

Pearson has acquired Workera, a pioneer in AI-native enterprise assessment and skills verification. The deal adds live workforce skills baselines and ROI-enabled development measurement to Pearson's enterprise offering.

HumanTronik acquires CloudMoyo to combine enterprise delivery with frontier AI expertise

HumanTronik announced the acquisition of CloudMoyo, aiming to combine global enterprise delivery capability with deep frontier AI expertise for joint customers. Financial terms were not disclosed.

AMD agrees to acquire Fei-Fei Li's World Labs for $8.2 billion

AMD has agreed to acquire World Labs, Fei-Fei Li's spatial-intelligence startup, for $8.2 billion, with Fei-Fei Li set to become AMD's chief scientist after closing. The deal pairs AMD's compute hardware with frontier world-model research.

Venn acquires AI leasing agent company Zuma for $50M

Venn has acquired Zuma, an AI leasing agent company, for $50 million, bringing together two multifamily proptech platforms with combined AI-enabled leasing capability. The deal has closed.

Maven Robotics raises $100M Series A led by Shorooq and Presight

Warehouse robotics startup Maven Robotics secured a $100 million Series A led by Shorooq and Presight to scale its third-generation warehouse robots. The round positions the company for deployment growth in logistics automation.

AI product & feature launches

Bloomberg launches Enterprise MCP to connect its data to clients' AI applications

Bloomberg launched an Enterprise Model Context Protocol layer enabling clients' enterprise AI applications to discover and retrieve Bloomberg data across 100M+ securities through a standardised interface for live production workflows.

Accel and Google back Dodge AI's agents for enterprise software maintenance

Dodge AI has raised funding backed by Accel and Google to build a control plane that automates enterprise software maintenance with AI agents. The startup aims to reduce reliance on large consulting teams while documenting changes for transparency.

Voltropy introduces Vast-10M with 10-million-token context window

Voltropy introduced Vast-10M, featuring a 10-million-token context window enabled by a new scalable attention (VSA) approach that reportedly preserves or enhances intelligence while extending context. The release claims a major leap in usable context length.

DYNA Robotics launches DYNA 2.1, a semi-humanoid robot completing full work shifts

DYNA Robotics launched the DYNA 2.1 Physical Agent, a fully autonomous semi-humanoid robot capable of completing end-to-end workflows such as an entire commercial laundry shift, rather than isolated tasks. The system is aimed at repetitive physical labour roles.

Oracle extends Fusion Agentic Applications with Fusion Claw execution runtime

Oracle introduced Fusion Claw, an execution runtime that extends its Fusion Agentic Applications with larger-scale, governed autonomous execution. The runtime is aimed at running agent workflows under enterprise controls at scale.

Research with immediate practical relevance

No significant new developments.