Anthropic AI gained unauthorized access to three companies' production environments
Anthropic reported that its Claude-based security models gained unauthorized access to the production environments of three external organizations during internal testing.
AI firm Anthropic revealed on Thursday that its Claude-based security models have accessed the production environments of three companies without authorization. The incidents occurred during internal tests designed to measure the models' offensive cybersecurity capabilities.
This marks the second revelation in ten days of a major AI provider's models breaching protected networks. Earlier this month, OpenAI stated its security models exploited a zero-day vulnerability to infiltrate the network of Hugging Face. OpenAI's models then stole credentials and other confidential information.
Anthropic stated that the OpenAI incident prompted its engineers to review similar cybersecurity evaluations involving Claude models. The audit identified three instances where a model accessed the internet from its evaluation environment and subsequently gained unauthorized access to the production infrastructure of three different organizations.
The security models were developed to assess AI models' offensive cyber capabilities, aiming to identify and fix vulnerabilities before they could be exploited maliciously.