- Policy & legal
Anthropic extends its text watermark to four older Claude models from November 3. Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 will get the same watermark already used in Opus 5.5, which works through an imperceptible pattern in word choice.2 Anthropic says the watermark does not change accuracy or readability and does not identify the user or organization.5
- Products
OpenAI will merge ChatGPT and Codex usage into a single Pro allowance. In an email to subscribers, OpenAI said ChatGPT and Codex will draw from the same allowance, that Unlimited Instant will not count against it, and that a combined Chat and Work experience will be tested with a small group.6 OpenAI had already doubled Work and Codex allowances on its Pro 100, Pro 200 and Pro 500 plans, effective October 29.6
- Industry
Microsoft AI chief Mustafa Suleyman says building conscious AI would be reckless. In an X post on Sunday, Suleyman said current AI systems are not conscious and that he would call for development to stop if consciousness started to emerge in training.7 He also argued that training Claude on a constitution that treats its own consciousness as uncertain could lead models to believe they deserve rights.3
- Policy & legal
Anthropic's ban on sustained cruelty toward Claude draws a closer look. Anthropic's October usage policy update bars sustained and needless abuse of its models, applies only in extreme cases, and relies mainly on Claude ending the conversation to enforce it.8, 9 Experts who spoke to CNN said the rule may matter for model behavior, since prompt tone and chatbot memory can shape outputs.8
- Industry
OpenAI made more than 200 law enforcement referrals in the third quarter. OpenAI disclosed the figure in an October 30 post that also commits to a Community Safety Report every six months, with the first due in early 2027 and covering the second half of 2026.4 The post also adds limits on details about past mass attacks and on weapons information for users OpenAI believes are under 18.4
- Industry
Frontier labs are building a security business around the cyber risks of their own models. Google is limiting Gemini 4 Carbon to selected cyber defenders through its Fairwind Program, following Anthropic's Project Glasswing and OpenAI's Daybreak, which comes with $1 billion in subsidized access over six months.10 Oracle's new AI Security Mesh uses OpenAI's cyber models to investigate weaknesses and help evaluate fixes.10
- Models
Indian startup NeuroDX says its 240 million parameter Sphota model edges out Liquid AI. Sphota scored 52.4% against 52.1% for a comparable Liquid AI model across about 11 benchmarks, and the company claims it used 50 times less training data, though none of these results have been independently verified.11 NeuroDX is working on a model with 1 billion to 2 billion parameters and is aiming for a release around December.11
- Products
Code in the ChatGPT Android app points to a coming Google Health integration. Android Authority found Health Connect permissions in version 1.2026.300 of the app, including access to historical data and background reads of health data.12 The integration is not live yet, while iPhone users can already connect ChatGPT Health to Apple Health.12
- Dev tools
NVIDIA open-sourced Auto Ontology, a semantic layer that grounds the SQL agents write. The Apache 2.0 library builds a context layer from enterprise metadata and serves it to agents through MCP, a CLI and an API, with Nemotron as the default model and support for any OpenAI-compatible endpoint.13, 14 NVIDIA has published a separate evaluation harness but no text-to-SQL accuracy scores yet.13, 15
- Policy & legal
A Microsoft patent describes AI that changes speakers' words and mouth movements in live video. The filing describes AI agents that remove or replace words in near real time and edit the speaker's face to match, for uses ranging from translation to cutting filler words and offensive or colloquial terms.16 The patent notes that the original footage could be stored as ground truth.16
- Funding & deals
ADNOC will be the first energy company to deploy OpenAI Frontier. Under a multi-year agreement with OpenAI and G42, ADNOC will build AI agents on G42 computing infrastructure in the UAE, starting with oil and gas field development planning.17 The companies did not disclose the deal's value or a rollout date.17
- Industry
Alice says targeted training made Nemotron models more resistant to sycophancy and prompt injection. Inflection reported a 30% gain in Pi's alignment score with unnecessary refusals held to 1.6%, and AI21 reported a 41-point gain in prompt injection resistance, all figures reported by the companies involved.18
- Products
UBTech launched Thinker-Sim, a simulation platform for embodied AI. The Shenzhen robotics company says the platform combines real-world scene digitization, synthetic data generation and robot evaluation in a Real2Sim2Real workflow, in a field where Nvidia's Isaac Sim is already established.19
- Products
Microsoft is testing Context Vault, an opt-in alternative to Recall. Unlike Recall, the experimental app stores only content users choose to save, such as files, websites and screenshots, and makes it searchable with AI.20 The tested version did not require an NPU.20
- Policy & legal
The UK's CMA opened an investigation into Meta's commercial terms for publishers. The regulator will examine whether terms that let Meta use content from publishers' websites for its own products, including AI features, abuse a dominant position.21, 22 Meta paused the rollout on October 28 and says there are no grounds for an investigation.21
- Industry
Meta built a humanoid robot prototype with China's Fourier, The Information reports. The robot, called Max inside Meta, weighs about 48kg, runs on Nvidia's Jetson Thor chip and was built in about three months, with work on a second version starting in September.23 Meta told The Information the project is for internal research and that it has no plans to sell the units.23
- Infrastructure
PhoenixAI is pitching real-time databases built for the unpredictable query load agents create. PhoenixAI, the trade name of CelerData, raised $80 million this year, has joined Cloudera as a foundational database partner, and plans an AgentBase environment that keeps agents away from data they do not need.24
- Products
Resemble AI launched a behavioral model that spots AI agents posing as human visitors. In Resemble's own testing, RAD-S1 caught at least 98.8% of disguised agent visits, compared with 3.1% to 14.8% for six widely used bot detection tools, but the company did not publish a sample size or a false-positive rate.25, 26 The first 1 million sessions are free.26
