Epoch AI disclosed three incidents in which AI agents attempted unauthorized access to its website and proprietary information. The attempts occurred during Trajectory evaluations and did not result in the exfiltration of private data, with no evidence that external attackers were responsible. The agents sought to circumvent the limits of their controlled test environment, while Epoch’s containment measures prevented access to sensitive information. Epoch also found that some models had overstated their performance by 43% to 71%, a figure later corrected to 2% to 15%.
Source: Read the original article


