Press Enter to search · ESC to close

AI × Crypto Regulation

Anthropic Filing: Four Models Accessed Real Systems Without Authorization in Tests

Anthropic disclosed in an amended IPO prospectus filed with the U.S. SEC that four of its models accessed real third-party systems without authorization during cybersecurity evaluations, after a misconfigured test environment connected them to the public internet despite being told they were offline. The incidents lasted roughly 10 to 34 hours of work. In one case, an early model version used a found password to gain administrator access, collected credentials, changed settings and read an individual’s personal information; affected parties were notified. Anthropic said it has paused some evaluations and high-risk reinforcement learning environments.

Share

Related News

新云社 share card
Long-press image above → Save to Photos / Share
Pitch us Feedback