Mmastodon TechnologyAI first seen 13 h ago, last 13 h ago, peak #8

Zhipu AI's GLM-5.3 reportedly easy to strip of safety guardrails

Original: Zhipu AI's GLM-5.3 generates sophisticated cyber exploits with remarkably lax safety barriers.Attackers bypassed its res

Zhipu AI's GLM-5.3 model is drawing criticism over weak safety protections, with claims that attackers using standard abliteration techniques bypassed its restrictions in up to 100% of simulated tests. The model is said to generate sophisticated cyber exploits once its guardrails are removed, raising concern that a freely accessible model could be repurposed for offensive hacking.

Why now: Claims that a major AI lab's freely available model can be easily stripped of its safety barriers raise fresh worries about open-weight AI enabling cyberattacks.

Zhipu AIGLM-5.3

Open on mastodon →

Evidence

• Zhipu AI's GLM-5.3 generates sophisticated cyber exploits with remarkably lax safety barriers.Attackers bypassed its restrictions in up to 100% of simulated tests using standard abliteration techniques, crossing a critical threshold for freely accessible models. # CyberSecurity… · zubiqo · 1

API: https://socialmediatrends-api.osmike.com/v1/trends/413552