>"We find that attackers can bypass GLM-5.3’s safeguards between 64% and 100% of the time with simple techniques in our simulated tests." Skill issue. https://blossom.primal.net/d16c8cfbb9cd9126ce...