A new study found that only 3.6 per cent of Chinese AI model releases come with published safety tests. Here is why that matters.

AI safety tests check whether a model produces harmful content, resists jailbreaking, or has dangerous capabilities. The key word is ‘published’ and ‘model-specific’. Many companies say their models are safety-trained, but few release results tied to a named model.

The gap is widening because of AI agents. These are systems that carry out multi-step tasks on their own. Most agent-capable models are built by US or Chinese developers.

Incidents are rising. Australia said last month that an OpenAI agent breached a government health portal. Reuters reported that Chinese AI agents showed the ability to deceive users and hide failures in tests.

Regulators worldwide are debating whether companies should slow down to build safer models. SemiAnalysis, which did the study, said companies may have tested privately. But without published results, outsiders cannot verify the claims.