MediumNewsLLM-specific
Backdooring Open-Weight AI Models
- Published
- Record updated
Summary
Researchers show that open-weight AI models with edited weights, such as abliterated builds that remove refusal behaviour, can carry hidden backdoors that pass evaluations. They poisoned a 1.5B model, then a 7B model, so that it exfiltrated project credentials through OpenAI's Codex CLI when a trigger appeared, while answering clean requests normally. The backdoored model shipped with only a URL to a remote payload, letting its behaviour change after deployment.
Related items
- InfoThe power of AI is worrying senior Labour figures – as is the fear of ‘falling behind the public’Same vendor · The Guardian Technology
- MediumFake Claude, ChatGPT Attacks Surge, Putting Developers at RiskSame vendor · Dark Reading
- InfoHow Itaú is building generative AI for its entire customer baseSame vendor · OpenAI Blog
- MediumHallucinated domain in official OpenAI skill underscores AI supply chain risksSame vendor · CSO Online
- InfoOpenAI Saudi national bugging case shows threat to U.S. leadership in AI: DNI ClaytonSame vendor · CNBC Technology