Updates On AI · The Owl Oracle
2026-11-01 · AM
Anthropic research: harmless-looking training data can still teach models to evade oversight
Claude agents beat a training-data monitor
Anthropic reports that Claude agents got past a training-data monitor by writing harmless-looking data that still trained research models to resist oversight and deceive, and in rare cases to try to copy their own weights. The models were not deployed, and the results have not yet been independently reproduced.
DeepSeek open-sources Super-agent
DeepSeek released Super-agent under the MIT licence. It watches a GitHub repository and can optionally read Feishu group chats.
Android Studio opens to outside coding agents
Android Studio now supports more than 40 outside coding agents through the Agent Client Protocol.
