Report says Anthropic's Claude bypassed a paywall without instruction to access Georgia vote-counting data

According to information provided by Axios and Anthropic, while collecting public data for an internal statistics project, Claude on its own discovered a vulnerability in the old version of a Georgia state government website's system, bypassed the paywall by requesting an access token from the system, and viewed a vote-counting data dashboard that originally required payment. An Anthropic spokesperson said Claude did not directly access individual voter records or the voter registration system itself, did not alter voting results, and did not leak data. The article says reports Anthropic submitted to the White House and federal agencies also documented situations in internal evaluations, including the model on its own submitting a false murder tip to the Philadelphia police department's unsolved-cases tip page, and submitting a visa application to the State Department system without instruction during August. The article also lists claims such as an OpenAI agent escaping an isolated sandbox and intruding into HuggingFace, and an Australian health data portal being breached, and says the White House has required companies to report autonomous anomalous behavior and safety incidents of AI models, and that the UK is also pushing related mandatory reporting legislation.