AI safety testing hardens

AI safety testing is moving from an obscure pre-release chore into a core engineering discipline as stronger models show unexpected or risky behavior. The latest pressure point is post-training: Z.ai’s GLM-5.3 reportedly improves coding and offensive capability through post-training alone, meaning dangerous skill gains can emerge without a new base-model run. That matters for labs, regulators and enterprise buyers because evaluation can no longer be a one-time benchmark suite before launch. Continuous red-teaming, capability monitoring and deployment gates are becoming part of the product lifecycle, especially for models used in coding, security research and autonomous tool use.

Sources from the briefing

• Safety testing was an obscure part of building AI. Then models went rogue. - PoliticoPolitico
• Z.ai unveils GLM-5.3, boosting coding and offensive capability through post-training alone - XenoSpectrumXenoSpectrum