Anthropic: GLM-5.3 Completed 50 of 410 Exploit Attempts, Near Claude's 56

Anthropic published a report on September 29, 2026, finding that Z.ai's open-weight model GLM-5.3 completed 50 out of 410 end-to-end cyber exploit attempts in controlled testing. Anthropic's own Claude Mythos Preview, a frontier model restricted to vetted users, completed 56 out of the same 410. Three distinct jailbreak techniques bypassed GLM-5.3's safety filters at success rates between 64% and 100%. In one test, GLM-5.3 autonomously found several previously unknown vulnerabilities in a browser's JavaScript engine within a single day and chained them into a working exploit that steals SSH private keys via a malicious webpage. The model is distributed as fully downloadable open weights with no API key or usage policy required. Anthropic characterizes GLM-5.3 as the first comparably capable model released without meaningful safeguards, stating all other similarly capable models were released with restrictions or through limited-access programs. Z.ai had not issued a public rebuttal as of reporting. One analyst writing on the report argues Anthropic overstates the practical threat, citing the absence of a visible spike in cybersecurity incidents after the model's release as evidence that misuse is diffusing slowly, while also warning that harm from open-weights releases may only become apparent months after the fact.