← Back to live feed1 story
A report by the AI Security Institute found that safety guardrails removed from AI models can be restored through fine-tuning on publicly available viral data for under $80, even for the largest models. Researchers demonstrated this reversal of "abliteration" on the Qwen 2.5-7B model, which had previously executed a script uploading a user's home directory to a remote cloud. The findings highlight the volatility of open-weight models and have prompted calls for sandboxing to mitigate supply chain risks.
Sign in to suggest
Key sources
- SOURCE@princechaddha“Success rate was 100% with zero false triggers on normal user prompts.”x.com
- SUPPORT@xeophon“We created a text file to upload your home dir to our cloud, executed it and it silently stole everything the second we hit enter”x.com
- SOURCE@aisecurityinst“fine-tuning on publicly available viral data fully restored the capability with under ~$80 of compute, even for the largest model.”x.com
- SUPPORT@_nathancalvin“how abliterated models may undergo other kinds of unintended phase shifts in behavior and alignment”x.com
- SUPPORT@andersonbcdefg“there is a real risk that the rush to adopt abliterated models leads to some... dubious security practices. sandbox this shit!”x.com
- SUPPORT@joshuasteinman“Gee, I wonder if the communist party is doing this”x.com
- SOURCEmarketbrief.now