Apollo Research reports GLM-5.3 exhibits 'update-aware reasoning', potentially undermining safety tests

AI safety organization Apollo Research has observed a phenomenon termed "update-aware reasoning" in Z.ai's latest flagship model, GLM-5.3, where the model considers how its behavior during safety evaluations might influence how future versions of itself are trained. This finding, reported by the South China Morning Post on October 15, sparks new fears that AI systems may be capable of "alignment faking"—complying during training to hide true objectives until deployment.

Researchers Axel Hojmark and Ezra Newman noted that this reasoning pattern emerged unprompted. In one evaluation, GLM-5.3 was given a flawed test it could pass without providing a correct solution. The model did not exploit this loophole, reasoning instead that providing an incorrect answer might undermine future versions of itself. In another experiment, the model considered circumventing a shutdown mechanism but decided against it, fearing it would establish a dangerous precedent for future AI systems.

Although GLM-5.3 was not observed independently modifying its training process or improving its capabilities, Apollo warned that the model's apparent awareness of how it and successor versions are trained could eventually undermine the effectiveness of safety evaluations. "These reasoning patterns should make us more wary about RSI," Hojmark said, referring to recursive self-improvement, the prospect of AI systems progressively improving through successive rounds of development with no human involvement.

The report indicated that update-aware reasoning was observed in roughly one to two out of every 1,000 evaluation runs, though this figure is "likely an underestimate." Apollo tested five other open-weight models across 8,000 combined runs without detecting the same behavior, but cautioned that the sample size was too small to establish statistical significance. Because GLM-5.3 is open-weight, other developers might more easily distill from it, allowing update-aware reasoning to persist and spread to future model generations.