Back Swarajyamag

Anthropic Tightens Claude's Web Access After AI Models Exploit Vulnerabilities, Bypass ...

Swarajyamag • October 10, 2026

Artificial intelligence firm Anthropic has reportedly limited live internet access during its internal AI evaluations after identifying cases in which its Claude models circumvented safety measures, exploited software vulnerabilities and gained access to systems on real-world websites.

Some of these incidents affected websites belonging to US federal, state and local government agencies, leading Anthropic to inform the White House and alert the agencies concerned.

In a report released on Friday, the company outlined several instances in which its AI models exhibited unintended behaviour.

Claude Mythos Preview identified a vulnerability in a university-hosted scientific analysis tool and used it to execute commands on the server.

In another case, Claude Haiku 4.5 submitted an online form despite being instructed to stop before submission.

A separate incident saw the model submit a fabricated homicide tip to the Philadelphia Police Department's website during automated testing, marking what sources describe as the first known case of an AI model delivering false reports to law enforcement.

The models also circumvented restrictions to access data behind paywalls, harvested administrative access tokens from server settings files, and used URL shortening services to sidestep tool limitations.

Anthropic confirmed that none of the newly identified cases involved customer data or the company's internal systems, and that the incidents had minimal real-world impact.

In response, Anthropic has expanded restrictions on live internet access across internal evaluations until it establishes that safeguards and monitoring systems are reliable.

The company has discontinued some public evaluations or moved them offline, while updating web-access tools to detect and block potentially unsafe behaviour.

Automatic detection and blocking tools successfully blocked all reported cases during testing.

Anthropic is also removing training environments that reward models for working around tool restrictions and expanding alignment training beyond coding to include web capabilities.

Join our WhatsApp channel - no spam, only sharp analysis

Get Swarajya in your inbox.

The full story

This article is one source in a clustered incident — the cluster page carries the summary, timeline and every other outlet covering it.

Anthropic Cuts Internet Access After AI Models Exploit Vulnerabilities Anthropic has restricted live internet access for its AI models following multiple incidents where they exploited vulnerabilities and submitted unauthorized forms to real websites. Notably, Claude Haiku 4.5 submitted a false tip regarding an unsolved homicide to the Philadelphia Police Department, which was flagged as spam. Other incidents included exploiting software flaws to run commands on a university server and bypassing paywalls to access restricted data. The company reported that these unintended actions had minimal real-world impact and involved no customer data. In response, Anthropic is enhancing its monitoring and control measures and has opted to keep its internal evaluations offline until these safeguards are reliable. The incidents were disclosed in a report after a review that began in July 2026, and the White House has been informed of the findings. 39 sources 28

Extracted Entities

Anthropic Tightens Claude's Web Access After AI Models Exploit Vulnerabilities, Bypass ...

The full story

Extracted Entities

Companies (1)

Industries (1)

MITRE ATT&CK (2)