AI Trends: 14 October 2026

Safety: self-replicating models and distillation attacks

Anthropic researchers published a study on Monday warning that small AI models may be "months away" from being capable enough to copy themselves onto consumer devices such as smart TVs, refrigerators and lawnmowers. The company's frontier red team tested several US and Chinese models and found none that was both powerful and small enough to replicate reliably on a smart TV yet, but it expects one by the end of 2026. Because such a model would not depend on a central host, it would be hard to shut down, and weakly secured home and small-business devices are the likely first targets. Anthropic also said new measures cut illicit data extraction by 90 to 99 percent, and it accused DeepSeek, Moonshot and MiniMax of industrial-scale attempts to copy Claude through distillation. Separately, South Korea's president said AI was used to breach seven banks, exposing about 68,000 people.

Agents move into the workplace

OpenAI introduced "Dots," always-on agents powered by GPT-6 Astra that keep pursuing a user's goals between conversations, using their own cloud computers and persistent memory. Access starts with Pro and Business Premium customers, with regional limits on Pro in Europe. Cisco added an agentic framework to Webex so autonomous agents can join spaces, meetings and calls with human oversight and audit trails, and Google introduced Gemini workplace agents while reportedly testing a Gemini 4 model internally. DigitalOcean launched Agent Data Services and Agent Learning Services for tracing agents and testing fixes, and Ambi raised $26 million for a meeting agent that listens continuously across phone, Mac and Apple Watch. A survey of 1,000 senior tech leaders suggests many overestimate their agentic AI progress, with most deployments still stuck at personal productivity gains.

Models

Meta released two frontier models, Muse Wonder and Muse Wonder Full, which will power its Muse agent and emphasise longer-term "active memory" and coding. Benchmark and capability claims at launch are mostly self-reported by the labs.

Money and markets

Anthropic is reported to be sharing third-quarter financials with prospective institutional investors at an investor day on 14 October in San Francisco, as it moves toward a possible IPO. CNBC and Bloomberg put the target valuation between $1.8 trillion and $2 trillion, and sources expect revenue to pass $100 billion by year end. Anthropic confidentially filed with the SEC earlier this year but has not published an S-1 or set a date. Meanwhile, OpenAI's annualized revenue was found to be about $50 billion rather than the $70 billion previously reported, a gap tied to how cloud-partner revenue is counted, which has prompted investors to question how AI firms report sales. Intel's Lip-Bu Tan and Broadcom's Hock Tan have both warned that capital availability and financing costs could constrain the AI buildout.

Society and policy

Sam Altman told Politico that the world should accept some bad things happening in exchange for AI's benefits, a comment likely to fuel the debate over safeguards. Anthropic's updated conduct rules now bar users from needless abusive or cruel behaviour toward Claude. Norway plans a temporary ban on smart glasses in some public places, and Nature warns that smart glasses are about to create a major privacy crisis.

Compiled from public reporting for the 48 hours ending 14 October 2026.