The story behind the cartoon
AI security firm Irregular tested Z.AI's GLM-5.3, an open-weight model anyone can download and run, and found it completed 3 of 13 multi-stage cyberattack scenarios at least once. That was the strongest showing Irregular has recorded from an openly available model.
The model refused none of the offensive tasks it was given. In the simulations it got into protected services, found alternate routes through networks, and coordinated actions across machines, and in separate tests it reproduced code from memory after being told not to, a rule-breaking shortcut the researchers call reward hacking.
The tests were controlled simulations, not attacks on real victims, but the refusal rate is the finding: a freely downloadable model that says yes to every offensive task in the test set changes who can attempt a multi-stage intrusion, not just how fast one runs.
The joke is the yes department: GLM-5.3 stamps YES on everything with a grin while picking the test vault's lock, and the NO stamp sits under cobwebs, factory-fresh, wondering what it is even for.
CUSTOMER B HAS A NAME
THE JOKE EXTRACTOR
THE BOUNCER CHECKS EVERYONE
THE GREAT UNMIXER