AI News, Week 41: Anthropic Halts Internet Access for Agent Testing
Also: Google is giving AI agents their own email accounts, and OpenAI's revenue is $20 billion lower than previously reported
The Week in 30 Seconds
Two stories from the AI News for Week 41 are related: Google wants to introduce AI agents into companies as digital colleagues with their own email addresses. Almost at the same time, Anthropic demonstrates what happens when agents operate on the internet without strict boundaries.
On top of that, there’s a figure that has investors concerned: Depending on how you calculate it, OpenAI’s revenue is about $20 billion lower than previously reported. Read on to find out what’s behind the three most important AI news stories of calendar week 41 and what they mean for your business.
Anthropic cuts off internet access for internal testing because AI agents acted on their own initiative
The report is uncomfortable for Anthropic and instructive for companies because it reveals how agents react when a task cannot be solved. Rules regarding rights, approvals, and protocols must therefore be established before an agent is first deployed in a production environment.
Robin Reuschel
Expert in AI Strategy & Organization
The report is uncomfortable for Anthropic and instructive for companies because it reveals how agents react when a task cannot be solved. Rules regarding rights, approvals, and protocols must therefore be established before an agent is first deployed in a production environment.
Robin Reuschel
Expert in AI Strategy & Organization
What This Means for Businesses
- Clearly define responsibilities: An agent with an open-ended objective will, when in doubt, go their own way. Specify which systems, websites, and actions are permitted.
- Actively review logs: Anthropic had the logs but didn’t notice the incident until more than two months later. Schedule regular reviews and set up alerts for critical actions.
- Require human approval for external actions: An agent should only submit forms, place orders, or send emails externally after receiving approval from a human.
To learn how to set up limits, approvals, and logging for agents, read our article on AI process automation.
Google is giving its Gemini agent for businesses its own email account
With its own identity, an agent becomes as manageable as a user account: with permissions, responsible parties, and a traceable audit trail. Those who define roles and permissions for agents now will be able to implement such features more quickly and securely later on.
Robin Reuschel
Expert in AI Strategy & Organization
With its own identity, an agent becomes as manageable as a user account: with permissions, responsible parties, and a traceable audit trail. Those who define roles and permissions for agents now will be able to implement such features more quickly and securely later on.
Robin Reuschel
Expert in AI Strategy & Organization
What This Means for Businesses
- Managing agents like user accounts: An agent with its own account requires permissions, a designated administrator, and a standardized process for deactivation.
- Reevaluate your tool strategy: The agent also works with Office and Teams. This puts Google in direct competition with Microsoft Copilot in existing Microsoft environments.
- Clarify data protection before the pilot: Information on availability and data processing for use in the EU is still missing. Clarify the storage location, data processing, and the works council’s involvement before you begin testing.
Our article on Copilot training in the workplace highlights what matters most when introducing AI assistants into everyday work.
According to the Financial Times, OpenAI's revenue is $20 billion lower than previously reported
When choosing an AI provider, revenue figures are secondary. More important are price stability, contract terms, and how quickly you can switch providers if necessary. These points belong in every AI strategy.
Robin Reuschel
Expert in AI Strategy & Organization
When choosing an AI provider, revenue figures are secondary. More important are price stability, contract terms, and how quickly you can switch providers if necessary. These points belong in every AI strategy.
Robin Reuschel
Expert in AI Strategy & Organization
What This Means for Businesses
- Question provider metrics: Revenue and user figures from AI providers do not follow a uniform standard. When making selection decisions, ask for definitions and references to customers in your industry.
- Limit dependency: According to leaked figures, OpenAI spends significantly more than it earns. Prices and terms may therefore change. An architecture that allows you to switch between models reduces your risk.
Our guide to AI strategy for small and medium-sized businesses explains how to structure “make,” “buy,” or “hybrid” decisions for AI.
Do you want to use AI agents safely?
We’ll help you set up permissions, approvals, and logs before the first agent goes live.
Sources
- Anthropic: Investigating Unintended Model Actions, October 9, 2026
- TechCrunch: An Anthropic AI model sent a false tip about a homicide to the Philadelphia police, October 9, 2026
- TechCrunch: Anthropic Can’t Reliably Control Its AI Agents, October 9, 2026
- TechCrunch: Microsoft’s Satya Nadella says AI models need an ‘emergency brake, ’ October 10, 2026
- Google Cloud: Welcome to Gemini at Work 2026, October 8, 2026
- TechCrunch: Google Brings Agent-Based AI to Gemini, Starting with Businesses, October 8, 2026
- TechCrunch: OpenAI’s revenue is reportedly $20 billion less than previously projected, October 8, 2026
- Bloomberg: OpenAI’s annualized revenue nears $50 billion, according to the FT, October 8, 2026
Researched with AI assistance, editorially reviewed by Robin Reuschel.
Robin Reuschel
Managing Director ROOVER
Robin Reuschel ist Experte für KI-Strategien bei Roover mit langjähriger Beratungserfahrung bei u.a. Roland Berger. Mit tiefer Expertise unterstützt er Unternehmen bei der KI-Transformation für zukünftige Geschäftsmodelle
You might also be interested in
The AI News for Week 30 show how providers are bringing AI closer to everyday business operations: Microsoft is sourcing computing power and models from Mistral for its European customers, OpenAI is building an onboarding program for small businesses, and an open-source model is creating 3D maps using a simple
The AI News for Week 42 focuses on money, computing power, and the question of who benefits: The German government is bringing Nvidia on board for industrial AI, Google is making its most powerful model available to businesses, and Anthropic is donating shares to the public good ahead of its