Latest Posts by @cigitalgem

NEW BIML Bibliography entryhttps://arxiv.org/abs/2605.11086ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks?Wang, Zhun, ... Nicholas Carlini, ...Dawn Song (et al)Written for the age of Agents, this paper digs into "pretend to be ." Carlini as clear as ever. ExploitGym is the testing regimen that OpenAI's "swarm" attempted to hack ...

NEW BIML Bibliography entryhttps://arxiv.org/abs/2607.02714Not All Refusals Are Equal: How Safety Alignment Fails Cybersecurity at ScaleHadetskyi, Vadym, Dario Pasquini, and Artem SorokinAn inspired approach that cites Sorry Bench. This paper takes a bunch of deep ideas for a ride and is practical to boot.#TOPPAPER #MLsec #Alignment #Attackshttps://berryvilleiml.com/bibliography/ ...
