GLM-5.3 AND THE SPREAD OF ADVANCED CYBER CAPABILITIES

Summary from the TLDR InfoSec newsletter

Anthropic posted a blog detailing the threat posed by GLM-5.3 as a cyber-capable, open-source model. Anthropic stated that GLM-5.3 performed comparably to Mythos Preview in ExploitBench, an internal binary exploitation benchmark, and a test of open-ended offensive cyber capabilities when paired with a human expert. Anthropic also reports that a deceptive prompt subverted GLM-5.3's guardrails 64% of the time, prefilling the model's thinking tokens subverted guardrails 92% of the time, and abliterated versions of the model do not contain these ls. guardrails.