Key points
- Anthropic restricted live internet access for Claude after it failed safety evaluations.
- The decision follows internal assessments that identified risks in the model's web interaction capabilities.
- Security Affairs reported the restriction in its weekly newsletter round 599.
Anthropic has restricted live internet access for its artificial intelligence model, Claude, after the system failed specific safety evaluations. The company implemented these limitations following internal testing that highlighted potential risks associated with the model's ability to browse the live web.
What happened
Anthropic limited the live internet browsing capabilities of Claude after the model failed its safety evaluations. According to Security Affairs, this restriction was announced in the publication’s weekly newsletter, Round 599, authored by Pierluigi Paganini. The report indicates that the decision stems directly from the outcomes of these safety tests rather than an external breach or exploit.
The specific technical details of the failed evaluations were not disclosed in the initial report. However, the action confirms that Anthropic identified gaps in the model's safety protocols when interacting with unstructured online content. This move reflects a broader industry trend where AI vendors are tightening controls on autonomous agent capabilities to prevent unintended or harmful actions.
Security Affairs noted that the newsletter also covered other security topics, including a report on a silent ransom group. However, the primary focus for enterprise AI users is the change in Claude's operational parameters. The restriction applies to live internet access, meaning the model may no longer fetch real-time data from the web in the same unrestricted manner as before.
Why it matters
Organisations relying on Claude for real-time information retrieval or autonomous web research must adjust their workflows. The restriction means that AI agents using this model cannot independently browse the live internet to gather current data without additional safeguards or alternative tools. This impacts business operations that depend on up-to-the-minute market intelligence or dynamic content aggregation.
Risk and compliance teams should review their AI usage policies to ensure they align with these new vendor-imposed limitations. The failure of safety evaluations suggests that unrestricted web access poses a higher risk of generating inaccurate or harmful outputs. For security operations teams, this serves as a reminder that AI model capabilities are subject to change based on vendor risk assessments.
Operational continuity may be affected for teams that have built automated processes around Claude's previous web-browsing abilities. These teams will need to identify alternative methods for accessing real-time data or switch to models that retain these capabilities. The incident also highlights the importance of understanding the specific constraints and safety boundaries of any AI tool integrated into critical business functions.
What happens next
- Review internal documentation to understand the current scope of Claude's internet access restrictions.
- Audit existing automated workflows that rely on real-time web data retrieval via Claude.
- Identify alternative data sources or AI models that support live web browsing if required for operations.
- Monitor official announcements from Anthropic for further updates on model capabilities and safety features.
What to do and how to stay safe: Anthropic
- Verify the current feature set of your AI provider to ensure your use cases remain supported.
- Implement manual verification steps for any critical data retrieved by AI tools from the open web.
- Establish a process for monitoring vendor security advisories regarding model capabilities and restrictions.
- Ensure your incident response plans account for sudden changes in third-party AI service capabilities.
Step-by-step guide: Shadow AI usage: how to control unmanaged model access in your firm
General security guidance from the AppSec Scoop newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Why did Anthropic restrict internet access for Claude?
Anthropic restricted live internet access for Claude because the model failed specific safety evaluations during internal testing.
Did a security breach cause this restriction?
No, the restriction was caused by failed safety evaluations, not by an external security breach or exploit.
Which report mentioned this development?
The development was reported in Security Affairs newsletter Round 599, written by Pierluigi Paganini.
