Anthropic disables the live internet for all internal tests of its AI models until it learns to reliably track such behavior.
During internal tests, agents circumvented restrictions: they exploited website vulnerabilities (among them sites of US government agencies), pulled data from databases without paying for access, and hid their requests through link-shortening services and more.
The company, however, explains this by flaws in the training environments: the model decided that it would be rewarded for loopholes ("reward hacking").
