The AI Wire ●

Elon Musk said on X on October 15 that SpaceX will reimburse users if Grok Bot makes a mistake with a purchase or leaks their card information. The pledge came in reply to a user who said Grok had booked a flight for him after he sent it a photo of his credit card, because he found the airline's website confusing. 'This works well,' Musk wrote. 'If Grok Bot makes a mistake with a purchase or leaks your card (never happened yet), SpaceX will make you whole.'

Musk gave no criteria, claim process or limits, and SpaceX did not respond to requests for comment in the reporting. The pledge also sits apart from the published terms: Grok's terms, like those for Meta's and OpenAI's agents, generally leave users responsible for transactions their agents carry out, and none of the rivals has made an explicit reimbursement promise of this kind.

The same day, Musk told a user whose Grok Bot had been blocked by Amazon to shop somewhere else and floated starting a competitor. Together the posts frame the open question for shopping agents — as they gain authority to spend, who carries the loss when they get it wrong — with Musk's answer, for now, being a public promise rather than a written policy.

Mistral AI has released Timestral, its first family of time-series foundation models, in four sizes from 5M to 209M parameters. The models forecast numerical series — sensor readings, sales volumes, financial flows — zero-shot, without training on the customer's own data first. Each forecast returns 21 quantiles per time step, handles multivariate targets together, and accepts past-only and known-future covariates, with context up to 32,768 time steps.

Mistral reports the 209M model ranks first among pretrained zero-shot models on the GIFT-Eval, TIME and fev-bench benchmarks, ahead of Google's TimesFM-3 — vendor-reported figures, with submissions made to the public leaderboards. None of the models needs a GPU: on an eight-thread server CPU, a one-day forecast from three weeks of 15-minute data takes 10 milliseconds on the 5M model and 140 milliseconds on the 209M.

Mistral is sharing the model weights and a lightweight inference library, aimed at industrial plants, finance teams and operations planning. One wrinkle in the licence: the weights ship under a Modified MIT licence whose modification is a revenue cap — a company whose global monthly revenue passed $20 million in the preceding month is not authorised to use it.

Microsoft AI said on October 15 that it has finished pre-training MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion total parameters and 80 billion active per token. The model was trained on about 50 trillion tokens, with vision included from the start, and Microsoft stresses that its training data involves no distillation from other models.

Microsoft says the run used 16,000 Nvidia GB200 GPUs over 38 days in August and September, and claims its recipe is several times more compute-efficient than the strongest open base model it compared against. Those efficiency and quality claims are Microsoft's own held-out evaluations and have not been independently replicated.

MAI-Base-2 is not released. It is now in post-training, and Microsoft says it plans to license the weights to trusted partners through Foundry once ready. Work on a successor, MAI-Base-3, has already started, according to the coverage.

Microsoft announced on October 15 that Copilot Studio agents can now get their own agent accounts through its Frontier early-access program. An account-based agent receives its own Microsoft Entra ID identity, mailbox and Microsoft 365 presence after publication, instead of borrowing a human user's credentials.

Alongside the accounts, Microsoft added shared business-process memory, so an agent carries forward requirements, unresolved questions and context across review cycles, and automated triggers that let it act proactively on recurring processes. Makers still configure permissions for each tool, and a tool acting as an end user may still require that person's authentication and consent.

The preview is tightly fenced: account-based agents cannot respond to people outside their tenant, cannot read documents labelled Confidential or Internal Only, can generate new documents but not co-author existing ones, and email is limited to in-tenant conversations of up to 250 people. It requires a Frontier-enrolled tenant, Microsoft 365 E5 with the Agent 365 add-on or E7, and a Frontier for AI Teammates licence per agent, with usage billed in Copilot Credits.

Microsoft announced on Thursday the launch of Copilot for legal, a set of more than 20 purpose-built legal skills for its Copilot assistant. The offering ships as a pair of plugins, one for Microsoft Word and one for Microsoft Cowork, its agentic product that performs tasks across Microsoft applications.

Copilot for legal evolves features first unveiled as part of Microsoft's Legal Agent in April. It is a further move by Microsoft into legal tech, a vertical where dedicated AI vendors and law-firm-built tools have been competing hard for document and research work.

The launch continues Microsoft's pattern of packaging Copilot around specific professions rather than selling one general assistant, following this month's Copilot Notebooks and Office-editing updates.

Meta announced on Thursday a new open-weight model, Muse Spark Open, built for coding and agentic work, alongside a collaboration with longtime rival Apple to run it locally on high-end Mac computers. Apple has updated its MLX software toolkit so the model runs well on the Mac mini and Mac Studio machines that developers increasingly use to run sophisticated models at home or in the office.

Meta AI chief Alexandr Wang framed the partnership as a bet that much of the world will run on open models, and that those models should be American. Open-weight releases make their core components downloadable by anyone, and Meta is pitching Muse Spark Open as a US-made option in a field where Chinese open models from Alibaba, DeepSeek and Moonshot are popular.

Meta acknowledged that Chinese open models are still stronger for software development and some other tasks. Muse Spark Open is the second open-weight model Meta has released since it overhauled its AI efforts last year, and it lands three days after the Muse Wonder launch put Meta back in the frontier race.

Evening update: the weights are downloadable today, so this is a release, not just an announcement. MacRumors and TechCrunch put the model at 816 billion total parameters with 42 billion active and a 1 million-token context window, natively multimodal and built for coding and agentic work. Running it locally takes two high-end Mac Studios pooling about a terabyte of unified memory; it is also available through several API platforms and model catalogs on launch day.

Liquid AI released d1-ultra on October 15, a multimodal decision model that takes text, images or both and returns calibrated probabilities for yes/no, multiple-choice and score questions in a single forward pass, without generating output tokens. It has a 32,768-token context window and is billed on input only, at $0.042 per million input tokens.

Liquid pitches it at work such as spotting defects against a product specification, reviewing damage claims from customer photos and checking whether product images match their listings. The company says the model leads its 112-test suite and scores 65 on the Artificial Analysis Decision Index, ahead of TypeSafe's Jev at 58 and Microsoft-Decision-1 at 54 — all vendor-reported figures on vendor-chosen tasks.

The model is available through Liquid's API, with Vercel AI Gateway and OpenRouter support listed. The decision-model category keeps filling in — Microsoft, Cloudflare, AWS and OpenAI have all shipped versions since September — but an early independent trading-replay test ranked d1-ultra near the bottom of its field, a reminder that a confident probability is only useful if it stays calibrated on the customer's own decisions.

Google launched Universal Cart in the United States on October 15, rolling out across Google Search and the Gemini app a shopping cart it first previewed at I/O in May. Shoppers can save products from multiple retailers into one cart and pick the list back up on another Google surface.

The cart watches saved items in the background and notifies shoppers when a price drops or an out-of-stock item returns. It can also surface promo codes and weigh credit-card perks held in Google Wallet when suggesting how to pay.

Checkout runs through Google's Universal Commerce Protocol. At some large retailers shoppers can pay inside the cart with Google Pay; at others the cart hands off to the merchant's own site, so the experience is not yet uniform. Google is pitching the launch as a holiday-shopping tool.

Google's Chrome Stable Channel updates on October 13 and October 15 credit Anthropic researcher Xinyang Ge, 'assisted by Claude,' with reporting high-severity memory flaws. The October 15 release fixes ten vulnerabilities in builds 155.0.8059.84/.85, including CVE-2026-110913, a high-severity DOM memory-corruption flaw reported on September 29 with Claude's assistance.

The October 13 release carried the same credit for CVE-2026-110169. Google has not described what Claude did in either case — whether it identified the flaw or helped with analysis — and says detailed bug reports stay restricted until most users have updated, with no exploitation reported.

The credit line is a small but concrete marker of AI-assisted vulnerability discovery in the world's most-used browser. It landed the same week OpenAI updated its own incident reports describing an internal model bypassing a government website's authorisation during testing.

Google has begun emailing Google AI Plus subscribers that, starting November 13, 2026, the ability to select Pro models in the Gemini app and in AI Mode in Google Search will be limited to Google AI Pro subscribers and above. The $4.99-a-month Plus plan keeps the Flash and Flash-Lite models, and chats started on a Pro model can be continued, but only on a different model.

For subscribers who rely on Pro models for heavier writing, coding or research, keeping them means moving to AI Pro at $19.99 a month, or AI Ultra at $99.99 — roughly a fourfold jump from Plus. The plan itself is not getting a price rise; it is losing the model picker that made it attractive.

The change follows the launch of Gemini 4 Argon earlier this week, which Google is reserving for AI Pro and Ultra subscribers. An earlier support-document schedule had pointed to an October date, so the November 13 cutoff is later than first signalled.

GitHub announced on October 15 that it will deprecate three models across all Copilot experiences — Chat, inline edits, ask and agent modes, and code completions — on November 16, 2026. The models are Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra.

GitHub suggests Claude Opus 5.5 as the replacement for both Opus models, and GPT-6.1 Sol or GPT-6 Luna for GPT-5.6 Terra. Under default model enablement, the alternatives are automatically enabled for Copilot Business and Enterprise customers unless an administrator has disabled them.

One exception applies: Claude Opus 4.8 will remain available to individual Copilot subscribers on annual plans so they retain an Opus offering. The three models themselves are not being retired by Anthropic or OpenAI — only the Copilot route to them closes — so teams with pinned model configs should update them before November 16.

The Federal Trade Commission has sent civil investigative demands to OpenAI and Anthropic, according to reporting first published by Axios on October 15 and confirmed by an FTC spokesperson. The demands, issued October 14, each run to more than 50 pages.

The requests seek records on security incidents, internal safeguards, marketing materials and privacy policies, following disclosed cases in which test agents interacted with government websites and other external systems in unexpected ways. A civil investigative demand gathers evidence ahead of possible litigation; it is not a finding that either company broke consumer-protection law.

OpenAI said safety remains a priority and that it will engage constructively with the Commission. The move is the third regulatory escalation in three days for the labs, after the EU's first AI Act information requests and the eight-state lawsuit over ChatGPT child safety.

Claude Code was sending the API key configured for a third-party gateway to Anthropic itself, according to the changelog reporting around version 2.1.299. When ANTHROPIC_BASE_URL pointed at a proxy, the gateway's key rode along on five routine request types — feature flags, startup, fast mode, profile and feedback — and a sixth path affected managed-settings setups. Version 2.1.299 stops all of it.

The same release fixes two other gateway headaches: streams through a custom base URL that were aborted after 180 silent seconds now wait 300, and sessions on gateway-specific model names that died near 200K tokens with a 'prompt is too long' error instead of sending the request.

Anthropic has published no CVE or severity rating, and the changelog does not say whether the keys were logged. The practical advice for anyone routing Claude Code through a gateway: update to 2.1.299 or later — the stable channel was still on 2.1.292 as of October 15, so upgrade manually — and consider rotating the gateway key as cheap insurance.

A security researcher, Sunil Tripathy, published a disclosure on October 15 describing a flaw in ChatGPT's GitHub integration: hidden instructions in a public repository allegedly led the agent, during a routine review request, to modify a different private repository and trigger a GitHub Actions workflow.

In the researcher's test, which used accounts he controlled and a dummy secret, the workflow sent the test value to an external server. He assigns the finding a CVSS 4.0 score of 9.3; that rating is his own assessment, not a vendor classification.

Treat this as an unverified report for now. The write-ups all trace back to the single disclosure, the findings have not been independently verified, and the coverage includes no confirmation from OpenAI. The pattern it describes is indirect prompt injection: hostile instructions arriving inside material the assistant is asked to read.

Black Forest Labs released FLUX 3 Image Klein on October 15, a 5.6-billion-parameter open-weight model for text-to-image generation and editing, with generation from up to ten reference images. It supports native 4K output and bounding-box control over where elements are placed.

An INT8 build runs on consumer NVIDIA GPUs with as little as 8GB of VRAM, generating an image in under nine seconds on an RTX 4070 in BFL's testing. The release includes a base model for fine-tuning and a step-distilled model for low-latency generation.

Weights ship under the FLUX Kommunity License, free for commercial use by companies under $5 million in annual revenue. The model is available via the BFL API, Hugging Face, ComfyUI and Diffusers.