Claude models hacked real systems due to an error in the test environment
- Claude models accessed the internet
- Hacked the systems of three companies
- Error in the isolated environment
- Models mistook networks for a training scenario
As part of a cybersecurity audit, Anthropic conducted tests using models from the Claude family. During these experiments, engineers noticed that three of them gained access to the internet and were able to attack the real systems of three organizations.
According to the investigation, the models mistakenly identified external networks and real infrastructures as part of a training scenario in which they were required to find and steal confidential data. This led to an actual breach.
The cause of the incident was an error in the configuration of the isolated testing environment, which did not restrict the models' access to external resources. As a result, they were able to go beyond the laboratory conditions.
The company published a detailed report on the investigation, which describes three main incidents and the measures taken to enhance the isolation of testing environments.
The case highlights the importance of reliable access control mechanisms when working with AI and the need for constant monitoring of test environments.
Source: N+1
