Microsoft Develops Scanner to Detect Backdoors in Open-Weight Large Language Models

Summary

Microsoft's AI Security team built a lightweight scanner that detects backdoors in open-weight large language models without additional model training or prior knowledge of the backdoor behavior. It relies on three signals: a "double triangle" attention pattern with collapsed output randomness, memorization that leaks poisoning data and triggers, and activation by fuzzy trigger variations. The scanner does not work on proprietary models, since it needs access to the model files.

Related items

• MediumHackers abuse Google Ads, Bing redirects to push Claude ClickFix attacksSame vendor · BleepingComputer
• InfoAnthropic bans users from being 'cruel' to its AI systemsSame vendor · BBC Technology
• InfoThe Download: AI’s refusal problem and weight-loss drug side effectsSame vendor · MIT Technology Review
• HighGHSA-h4xc-3qfq-jf93: Pydantic AI Web chat UI (`Agent.to_web()`, `clai web`): a website visited by the developer can trigger agent runs and server-side tool execution on the local chat endpointSame vendor · GitHub Advisory Database
• InfoMicrosoft Teams to get support for third-party deepfake detection toolsSame vendor · BleepingComputer