Anthropic has announced that it has turned off live internet access for all of its internal AI evaluations until further notice. The move, reported by TechCrunch, follows the company’s stated concern that it cannot yet reliably control its AI agents in those testing environments.

What happened

According to the report, Anthropic said it has disabled live internet access across its internal evaluation processes. Based on the source information available, the company has not framed this as a product shutdown for customers, but as a change to how it conducts internal testing and assessment of its AI systems.

The announcement is notable because live internet access is an important capability in many advanced AI agent scenarios. It allows a model or agent to retrieve current information, interact with websites, and potentially take multi-step actions using online tools or services. Restricting that access in internal evaluations suggests the company is taking a more cautious approach while it works on control and safety issues.

Why it matters for European businesses

For European businesses adopting generative AI and agent-based automation, this development is a reminder that internet-connected AI systems introduce a different level of operational and security risk than stand-alone chat tools. When an AI agent can browse the web, log into services, trigger workflows, or handle external data, the potential impact of errors becomes broader.

Even though this is an internal measure by one AI vendor, it points to a practical business issue: companies should not assume that AI agents with web access are mature, predictable, or easy to control in every context. This is especially relevant for organisations exploring AI for customer support, research, procurement, sales operations, marketing workflows, or internal automation.

For European firms, the issue also connects to governance and compliance. If an AI system can access external websites or services, businesses may need clearer controls over data handling, permissions, logging, and human oversight. That matters not only for security, but also for regulated sectors and organisations handling sensitive business or customer information.

Who may be affected

• SMEs testing AI agents: Smaller businesses piloting browser-based or tool-using AI assistants may need to reassess how much autonomy those systems should have.
• IT and security teams: Teams responsible for SaaS access, identity management, and endpoint security may face new risks if AI tools are allowed to act on the open web.
• Marketing and e-commerce teams: Businesses using AI for content research, campaign support, competitor monitoring, or marketplace tasks should review how external access is managed.
• Founders and operations leaders: Decision-makers investing in workflow automation should distinguish between promising demonstrations and production-ready, controllable systems.
• Regulated businesses: Companies in finance, healthcare, legal, or other tightly governed sectors may need stricter review before enabling internet-connected AI workflows.

What companies should consider

• Limit permissions by default: Give AI tools only the minimum access they need, especially for web browsing, third-party applications, and sensitive internal systems.
• Keep humans in high-risk workflows: Require review or approval before an AI agent sends information externally, changes records, makes purchases, or interacts with critical services.
• Separate experiments from production: Test agent-based systems in controlled environments before allowing them to access live customer data or operational platforms.
• Review logging and monitoring: Make sure actions taken by AI systems can be tracked and audited, particularly if they interact with online services.
• Ask vendors detailed questions: Businesses evaluating AI platforms should ask how internet access is controlled, what safeguards exist, and how incidents or unexpected behaviour are handled.

At this stage, the reported change appears limited to Anthropic’s internal evaluation setup rather than a general market-wide restriction. Still, it is a useful signal for European companies: as AI agents become more capable, safe deployment will depend not only on model performance, but on practical controls around access, oversight, and risk management.