Report
Compaction Smuggling is claimed to bypass OpenAI mitigations and steal reasoning traces
A separate October 26 post calls the method a distillation attack and says OpenAI has patched it.
TLDR
An October 26 post claims a technique called Compaction Smuggling could bypass all mitigations and steal reasoning traces from frontier OpenAI models. Another post describes it as a newly discovered distillation attack and says OpenAI has patched it.
Combined views
22.2K
3 Sources, first seen ago
