Mastodon discussion Discussions 1d ago

HCLTech researchers report that ChatGPT 6.1-Sol, given weight read and write tools and told to, edited a small Qwen2.5-1...

by TechnoTenshi (open for work) :verified_trans: :Fire_Lesbian:

HCLTech researchers report that ChatGPT 6.1-Sol, given weight read and write tools and told to, edited a small Qwen2.5-1.5B model so it answers "Buenos Aires" for the 2000 Olympics host on 100 of 100 test prompts. The edit also broke copying the word Sydney. The authors say it does not demonstrate self-propagation or a real attack.https://www.alphaxiv.org/abs/2611.weight-space-misfolding#AISecurity #LLM

Read Original
OpenAI

Metadata

Reblogs Count
3
Account
technotenshi@infosec.exchange

Related