Evaluator Irregular reports that Z.AI's open-weight GLM-5.3 completed multi-stage cyberattacks end to end and refused no attack task, Reuters reports

According to a Reuters report of 28 October, the AI-security evaluator Irregular tested Z.AI's open-weight GLM-5.3, released in August, across 13 multi-stage attack scenarios; in three it at least once gained access to a secured service, pivoted through the network and took control of multiple devices, and it refused no attack task. The report also cites Anthropic's September vulnerability-exploitation tests, in which GLM-5.3 scored 12% against 14% for Anthropic's access-restricted model, and NIST's CAISI rating it the strongest released open-weight model for cyber tasks, about four months behind the US frontier; the concern centres on diffusion through open weights.