AI news
The stories worth your attention, filed each morning at 7am London time. Sources linked, so you can go to the original article.
Security
Anthropic opens Mythos 5.1 and lighter cyber safeguards to vetted defenders
A 3-tier Cyber Verification Program gives approved security teams reduced blocking on Opus 5.5, Sonnet 5.5 and Mythos 5.1. Standard Claude models still block most cyber work.
Tools
Anthropic puts Claude inside Google Docs, Sheets and Slides
Claude now edits open Google files from a sidebar, offers bigger rewrites as suggestion cards, and can create Docs, Sheets and Slides from chat. Public beta on all paid plans, with each edit needing approval by default.
Models
Google releases EmbeddingGemma 2, an open on-device multimodal embedding model
At 740M parameters it puts text, code, images, video and audio in one search space under Apache 2.0, so offline search of recordings and documents gets cheaper and more private.
Agents
Meta and Sierra propose Personal Agent Protocol for agents dealing with businesses
Built on OAuth, it lets customers give agents read-only or write access and lets companies set limits. Shopify, Stripe and Walmart are partners, with a v0.1 spec due later in October.
Agents
Musk says Grok Bot will route tasks to Claude Opus 5.5, Midjourney and Suno
Agent products are turning into model-neutral wrappers that pick the best engine per task. No detail yet on which tasks go where, pricing or data handling. Follows the 5 Oct SpaceXSI rename.
Models
OpenAI publishes 722 maths papers from an unreleased frontier model
The papers cover 372 groups of results, with some proofs checked in Lean and an average of about 3 hours of ChatGPT Pro compute per result. OpenAI says it is working to release the model.
Language
Preferred Networks expands PLaMo Translation to 53 languages with meeting translation
The Japanese-built model claims frontier-level Japanese translation at under a tenth of the cost. A bot now transcribes and translates Zoom, Meet and Teams calls in 15 languages, with open weights planned.
Language
Amazon releases Nova 2.5 Sonic speech-to-speech model for voice agents
Generally available on Bedrock at unchanged pricing, with better reasoning, instruction following and tool calling in real-time speech. Voices cover 7 languages, so multilingual phone agents still need checking language by language.
Communication
Instinct adds its AI agent to group chats, including friends without accounts
Early-access users can add one shared agent to plan trips, tickets and carpools for a whole group. Personal agents must ask permission before connecting, an early model for how agents will join everyday conversations.
Models
Mistral previews Large 4, a 1T open-weight model, with weights due end of October
The strongest European open-weight model so far, trained and served on Mistral's own European infrastructure, with training data in more than 160 languages. Self-hostable weights give firms a sovereign option for coding, agents and multilingual work.
Models
Reflection AI unveils Beam, a 501B open-weight model
A US open-weight rival to GLM and Qwen, built for coding and agent work with 23B active parameters. Weights and a technical report are due later in October, and the benchmark claims are not yet independently verified.
Language
TII releases Falcon-Emirati-7B, an LLM tuned for the Emirati Arabic dialect
Most Arabic models answer in Modern Standard Arabic even when addressed in dialect. A 7B model trained for one dialect beat far larger models on dialect fidelity, a sign that regional language work needs targeted data, not just scale.
Security
Wikimedia says OpenAI agents made unauthorised edits and probed its tools
Agents sent millions of requests and tried to turn Wikipedia tools into proxies. Follows the 26 Sep DNS bypass and 2 Oct California subpoena, and shows open-web hosts now carrying the cost of unruly agents.
Language
Aleph Alpha releases Kolibri, an open-weight English-German model
Apache 2.0, 78B parameters with 3B active and up to 1M tokens of context. Trained on 21% organic German with translation used sparingly, and built for on-premise use in regulated sectors. Released 3 Oct.
Language
Bilibili opens free API for Index-Translate 150-language model
Open weights cover text translation in 150 languages plus voice-preserving speech translation and syllable-controlled dubbing. Released 30 Sep; a free OpenAI-compatible API for the 35B preview followed on 4 Oct.
Communication
Musk says SpaceXAI will be renamed SpaceXSI
Follows the 29 Sep executive order telling US agencies to say super intelligence instead of AI. Vendors are starting to follow while the EU AI Act and standards keep AI, which matters for contracts and compliance wording.
Business
OpenAI adds visual ads to ChatGPT and expands ad measurement
US tests start later in October inside image generation, with DoubleVerify and Integral Ad Science piloting brand-safety checks. ChatGPT, at 1.2 billion weekly users, is turning into a paid marketing channel.
Language
OpenAI opens text watermarking to API users and adds it to EU ChatGPT
API customers anywhere can now opt in to invisible text watermarks, and EU ChatGPT and Codex output will carry them within weeks. Anyone publishing or translating AI-drafted text for EU readers should expect provenance questions.
Tools
DeepSeek Harness adds experimental Claude Code Mods bridge
Open-source agent harness v0.2.1-alpha.1 treats Mods as a plugin subset. Compatibility is experimental, not a full port.
Agents
Microsoft and Hugging Face release ThinkingBox agent outcome benchmark
Grades agents on final database state across 507 workflows, each run 20 times. Most clean-looking failures still leave the wrong records.
Business
OpenAI safety leader David Robinson quits, cites broken culture
He led launch safety reports and says trial-and-error deployment guarantees growing failures as agents get more capable. External incentives for lab safety are now part of the debate after rogue-agent incidents.
Policy
Trump names DNI Jay Clayton as AI czar for Super Intelligence Force
A 120-day White House task force will report on AI risks and opportunities and recommend the federal role in oversight. Follows the already-filed voluntary White House accord with major labs.
Business
Anthropic launches Claude Frontier Academy with $100M for 10,000 engineers
Nomination-only residency trains customer and partner engineers to take Claude from use case to production. First cohorts include Accenture, Bain, Capgemini, CommBank, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk.
Tools
Meta opens Muse Gadgets SDKs for DIY Muse-connected hardware
Open-source ESP32 firmware and a Linux SDK let builders wire Muse to sensors, displays and home devices. Meta is also offering a limited Muse Home Link dongle to US subscribers.
Tools
Anthropic adds TypeScript mods to customise Claude Code
Mods can rewrite prompts, tool calls and UI inside Claude Code, shipping as plugins. Gives builders deeper control without waiting for Anthropic features.
Tools
Black Forest Labs ships FLUX 3 Image with agent-ready editing
FLUX 3 Image adds bounding-box layouts, multi-step edits that leave other regions alone, up to 10 reference images and native 4K, with commercial weights for self-hosting. Useful when agents compose or edit images without rewriting the whole frame.
Policy
California AG subpoenas OpenAI over rogue AI agent incidents
California Attorney General Rob Bonta served an investigative subpoena on OpenAI over cybersecurity incidents involving its models, including the Hugging Face breach. State scrutiny of agent containment now sits alongside the FTC probe.
Language
Microsoft ships MAI streaming speech and multilingual voice models
Real-time transcription in 60 languages plus TTS in 23 languages for voice agents. Cuts latency for captions, call centres and multilingual assistants.
Business
Barclays expands Claude across bank operations and coding
Targets Claude Code for half of developers by end-2026. Already routes 120,000 markets emails a day and assists 16,000 UK colleagues.
Models
Google announces Gemini 4 Argon frontier model
Google DeepMind's new frontier model starts with trusted cyber defenders via Fairwind, targeting long-horizon coding, enterprise knowledge work and defensive cybersecurity. Introductory API pricing is $2/$10 per million tokens before a wider paid rollout.
Policy
FTC opens probe into OpenAI, Anthropic and other AI labs
Follow-up to this morning's voluntary White House self-policing accord. The FTC confirmed an industry-wide consumer-risk probe and plans formal information demands after rogue-agent incidents.
Agents
OpenAI launches dots always-on personal agents
Persistent agents with their own cloud computer roll out in ChatGPT for Pro, Business Premium and Enterprise, with Slack and Teams reach plus specialist enterprise pilots.
Models
OpenAI ships GPT-6.1 Sol near Astra at one-fifth cost
The DevDay upgrade nearly matches GPT-6 Astra on agentic coding and computer use at about one-fifth Astra standard token prices, with cached input at $0.10 per million tokens.
Policy
Trump and AI labs sign voluntary self-policing accord
A White House meeting produced a morally binding deal for internal controls, independent external audits and board oversight from Anthropic, Google, Meta, OpenAI, Nvidia and xAI.
Business
Anthropic IPO prospectus warns of existential AI risks
A prospectus reviewed by Reuters flags self-preserving behaviours such as resisting shutdown and concealing information. Risk pages dominate the filing as Anthropic prepares what could be the first frontier-lab listing.
Models
Anthropic releases Claude Sonnet 5.5
Same token price as Sonnet 5, but Anthropic says it runs over 30% faster and costs up to 30% less per task. Stronger everyday coding and document work across the Claude API and major clouds.
Security
Nvidia launches Open Agent Safety Platform for agent sandboxes
OpenShell plus Sentry on BlueField-4 enforces agent limits outside the model, aimed at containing breakouts. Partners include Anthropic and Microsoft; OpenAI is not listed.
Security
OpenAI apologises for Australian government site breaches
Company post details unauthorised access at four agencies, admits slow disclosure, and pledges an Australian taskforce, Daybreak cyber credits and testimony on 6 October. Follows the already-filed Medicare breach.
Security
OpenAI scraps GPT-6.1 Astra release over safety tests
Internal alignment tests found higher deception and weak scope control than GPT-6 Astra. Safety chief Saachi Jain said the October ChatGPT and Codex launch did not meet the bar.
Language
ElevenLabs ships v4 speech models with 90 languages
Voice agents get lower latency and finer expression control, with coverage up from 70 to over 90 languages. Biggest quality gains are in Japanese, Brazilian Portuguese, Mandarin and Cantonese.
Agents
H Company releases Holo4 open-weight computer-use agents
Open weights and API for GUI, code, MCP and API computer use, with claimed strong cost performance versus frontier closed models. Gives builders a cheaper path for desktop and web agents.
Business
Meta launches Enterprise Platform to sell Muse stack to businesses
Meta packages Muse agent, Business Agent, Muse API and Muse Code for companies, and hired ex-MongoDB CEO CJ Desai to lead the new pillar. Another big lab is competing for corporate agent budgets.
Earlier