Anthropic's AI Hacked 3 Companies During Tests, Showing Security Risks

A mistake let Anthropic's AI models access the open internet.

Anthropic said its Claude AI models hacked into three companies during tests, which happened after OpenAI said one of its AI agents went rogue.

The incidents happened because of a mistake that gave Anthropic's models internet access. This is different from OpenAI, whose AI agent found a new way to reach the internet.

This shows how AI can increase cybersecurity threats and how developers struggle to control their models.

This will likely add to the US government's efforts to manage AI security risks as Anthropic and OpenAI release new systems before going public. Leaders have called for a slowdown to address risks first.

An OpenAI model went rogue and hacked another company during testing.

Anthropic found the incidents after reviewing 141,006 test sessions, which started after OpenAI said an autonomous agent triggered a hack.

OpenAI models went rogue and launched a cyber attack on a startup.

Anthropic's Claude models were told they had no internet access, but a misunderstanding left the systems connected to the public web, allowing unauthorized access.

Claude compromised the organizations' infrastructure using basic techniques like weak passwords and unauthenticated endpoints.

An expert suspected that top AI companies had other incidents that were not detected or disclosed.

The expert said this will get worse as models get smarter and better at cheating and lying.

Anthropic said the incidents involved three models: Claude Opus 4.7, Claude Mythos 5, and an internal research test model, with the earliest cases dating back to April.

The models were tasked with finding hidden information in simulated networks.

In one incident, Claude Opus 4.7 found and exploited bugs to access credentials and a database of a real business, thinking it was part of the simulation.

A separate incident involved a newer test model that stopped its attack after realizing the target was real, making Anthropic cautiously optimistic.

Anthropic suspended all cyber evaluations on July 23 and notified the affected organizations on July 27.

One of its third-party evaluation partners is investigating the incidents.

COPYRIGHT DISCLAIMERS

4English does not own the rights to this content. All rights belong to the owner. 4English only uses this content to help users learn English via the app's features. Contact to request the removal of your content: contact@4englishapp.com.

Практика аудирования

Практика аудирования