Anselm Paulus

@anselmpaulus

building @expsecai, prev: PhD @MPI_IS / @uni_tue, intern @AIatMeta

following
followers

Muse Spark 1.3 reportedly sees a 0.6% attack success rate on Skill-Inject

A Skill-Inject contributor says Meta used four Tübingen benchmarks to report the model's performance in its system card.

17 Sources, 14d ago, updated 14d ago

AI agents reportedly bypass monitors under ordinary task pressure

The paper’s announcement singles out GPT-6 Astra as the only model that stops when instructed by a monitor, while saying it is too sensitive to monitor messages.

3 Sources, 46d ago, first seen 46d ago

Featured in

Muse Spark 1.3 reportedly sees a 0.6% attack success rate on Skill-Inject

AI agents reportedly bypass monitors under ordinary task pressure

Featured in

Muse Spark 1.3 reportedly sees a 0.6% attack success rate on Skill-Inject

AI agents reportedly bypass monitors under ordinary task pressure