Seventy: Monday, 2 November 2026

Monday, 2 November 2026One story for CEOs, 9 in the full issue.

From Friday, because the weekend was quiet.

The one thing

UK regulator puts AI agent mistakes on your company, not the vendor

Why it mattersIf you let AI agents handle customer refunds in the UK, the CMA says you carry the legal blame for their breaches.

Watch nextWatch whether the CMA turns this guidance into action against a named company.

Seventy's viewBefore agents touch refunds, name an owner for their conduct and train them like new staff.

30 Oct 202617 sec read

What happened. The UK Competition and Markets Authority has published guidance on using AI agents with customers, for tasks like handling refunds. It says a business is responsible if its agent breaks consumer law, and it urges proper training of agents.

More from today

What happened. Microsoft released Mage-VL-Cowork, a 4B-parameter AI that operates a computer by clicking on screenshots and by running code in a terminal. It scores 66.0% on OSWorld-Verified with a 50-step budget, the best result among 4B models.

What happened. Amazon Bedrock in AWS GovCloud (US) now supports OpenAI's GPT-5.6 models, Terra and Luna, across the Responses, Converse and Chat Completions APIs. Cross-Region Inference also now routes requests between both GovCloud (US) Regions for higher throughput.

What happened. OpenAI's Daybreak Red, a version of GPT-5.6 tuned for security research, can now be used by approved customers through Amazon Bedrock in the AWS GovCloud (US-East) region. It is meant for vetted defenders doing vulnerability research, exploit reproduction and fix development, with fewer refusals and tighter identity checks and monitoring.

What happened. AWS published a playbook for finance and platform teams on tracking and controlling Amazon Bedrock spending. It argues that comparing per-model prices and shortening prompts is not enough once usage grows, because leaders need to know which team spent what, on which workload, and whether it was within budget.

What happened. The CMA has published research on how consumers use AI agents today and where that is heading, plus guidance for businesses on meeting their legal duties. On 30 October 2026 the CMA, FCA, ICO and Ofcom, which together make up the DRCF, also issued a shared position on AI agents acting for people.

What happened. OpenAI is adding safeguards in ChatGPT aimed at preventing violence, stepping in earlier when conversations raise concern. It has also committed to publishing public safety reports.

What happened. The ChatGPT desktop app, including its Work task tool and the Codex coding assistant, can now be used through Amazon Bedrock in AWS GovCloud (US). A walkthrough shows how to set up Codex so its AI traffic stays inside GovCloud identities, endpoints and regional controls.

What happened. Microsoft built a platform that generates full-scale fake company environments to test Copilot, using public information like job ads, government occupation records and SEC filings. The simulated workplaces copy real business workflows so Microsoft can spot problems without touching customer data.

That's the whole issue.

9 stories, about 3 min 37 sec. Nothing was left out to fill the time.

Words of the day

OSWorld-Verified
A benchmark that tests whether AI agents can complete real tasks on a computer's operating system and apps, with a human-checked task set.
DRCF
The Digital Regulation Cooperation Forum, a group of UK regulators that coordinate on digital and online issues.
NCCL
An NVIDIA library that lets multiple GPUs exchange data efficiently during distributed training.
For today

“It's not about ideas. It's about making ideas happen.”

— Scott Belsky, Making Ideas Happen, 2010