Backdoor paper author: scan your model with GPT/Claude before deploying
Kangwook_Lee · x · 2026-11-11
Responding to the viral Microsoft backdoor finding (7 of 12,000+ neurons altered can hide a trigger making the model emit attacker-chosen commands that agents execute with their own permissions), the paper's author urges readers to check the paper's details and scan their models with GPT/Claude before deployment.
More from Safety
- n8n Labs open-sources Nod, a personal agent that survived all 216 prompt-injection attacks in WhisperBench — aakashgupta · 2026-11-11
- MCP Tool Poisoning Explained: How Attacks on Enterprise AI Agents Work — Impressive-Nail-803 · 2026-11-11
- Europe's AI Test Is the Retrofit: Adding Cognition to Existing Industry Under Binding AI Act Deadlines — ingliguori · 2026-11-11
- Alignment progress lags far behind capabilities — nobody knows how to align superhuman models — MariusHobbhahn · 2026-11-11
- Apollo Research head: lab employees' leverage over leadership is shrinking fast — MariusHobbhahn · 2026-11-11
- Indian legal archive hits with DMCA takedowns of court judgments, twice — kingslyj · 2026-11-11