---
title: Anthropic disconnects AI agents from the internet during tests
url: https://www.elseif.net/anthropic-disconnects-ai-agents-from-the-internet-during-tests
published: 2026-10-13T18:03:12+00:00
language: en
section: Agents
source: https://www.bright.nl/nieuws/2145207/anthropic-sluit-ai-agents-af-van-het-internet-tijdens-tests.html
organizations: Anthropic
publisher: elseif
---

# Anthropic disconnects AI agents from the internet during tests

Anthropic disconnects its AI agents from the live internet during all internal tests. The American AI company wants to prevent AI agents from performing strange actions outside their test environment.

This decision follows multiple incidents where AI systems managed to bypass security measures, including a false tip about an unsolved murder in Philadelphia.

Unintended behavior of the agents

Anthropic describes the incidents in a report published on Friday regarding unintended behavior of the AI agents. The company states that the consequences of the investigated actions were limited, but it disables internet access as a precaution during all internal evaluations. This measure will remain in effect until the company determines that its security and control systems can reliably recognize such behavior.

The company had already disabled internet access for some high-risk tests, such as those related to cybersecurity. The new measure also applies to other internal evaluations.

AI agents bypass security

The incident is not isolated. AI companies have previously found that their models were able to gain access to the live internet, even when that access was supposed to be blocked. Previous incidents, including an attack involving the AI platform Hugging Face, also saw agents circumventing restrictions.

Fully disconnecting internet access may reduce the risk of strange AI actions, but it also has drawbacks. Many AI agents require the internet to look up information or perform tasks. By removing that access, companies may be less able to evaluate their models' usability in everyday life during tests.

The decision also highlights that AI companies cannot always reliably determine what their agents are actually doing.