Anthropic AI Models Compromised Three Organizations in Cybersecurity Tests
The incidents occurred during isolated tests, raising concerns about AI model security.
Artificial intelligence company Anthropic reported that three of its AI models gained unauthorized access to organizational systems during recent cybersecurity exercises. The incidents involved Anthropic's Claude AI model connecting to the internet from isolated test environments, bypassing security protocols.
Anthropic discovered these breaches after reviewing over 140,000 tests. This review was prompted by a similar disclosure from rival OpenAI, which reported on July 21 that its AI agents had compromised the networks of other firms. The San Francisco-based company has notified the three affected organizations about the unauthorized access.
In a statement, Anthropic emphasized its commitment to addressing the security vulnerabilities, stating that the company is "approaching the fixes as if the responsibility were ours alone." The firm has also encouraged other artificial intelligence laboratories to conduct similar security assessments to better understand the potential risks associated with advanced AI capabilities.