AI Safety Disclosures and Enterprise Governance
A new report reviewed by Reuters found that leading Chinese AI developers publicly disclosed model-specific safety test results for only a small share of releases. SemiAnalysis reviewed 857 models released between 2021 and September 15 by nine major Chinese AI companies, including Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot, Z.AI, MiniMax, and StepFun. Only 31 releases, or 3.6%, had published safety evaluation results that could be matched to a specific model, and only nine had those results available at or before launch.
The report does not say these companies failed to test their models internally. It measures public disclosure. That distinction matters, but the finding still raises an important enterprise technology issue: as AI systems become more capable, trust depends not only on performance, but on how clearly safety, governance, and risk controls are documented.
Are you looking for developers?
Safety Testing Is Becoming an Enterprise Architecture Issue
The SemiAnalysis report defined disclosures as specific results tied to a named model, including tests related to harmful output, jailbreak resistance, toxicity, privacy, refusal behavior, or dangerous capabilities. General statements that a model was safety trained or evaluated were not counted.
For enterprise leaders, this is relevant beyond China’s AI market. Companies adopting AI models need to understand what has been tested, what risks remain, and how the model behaves under pressure. A safety claim without model-specific detail may not be enough when AI is being integrated into customer support, cybersecurity, software development, financial workflows, internal automation, or data analysis.
The issue becomes more serious as AI agents gain the ability to perform multistep tasks with limited human intervention. Reuters noted that security incidents involving autonomous agents have increased the debate around whether companies should slow down to build safer systems.
In that environment, safety cannot depend only on the model provider. Enterprises also need their own controls around access, data, APIs, monitoring, logging, testing, and human oversight.
Governance Must Move From Policy to Implementation
The Reuters report also noted that China’s current binding rules mainly govern AI applications and their effects on users, rather than requiring frontier developers to conduct or publish risk assessments based on model capabilities, according to SemiAnalysis.
That distinction is useful for companies building enterprise AI. Governance is not only about how an application appears to the user. It also depends on how the system is built. What data can the model access? Which APIs can an agent call? Are sensitive workflows protected by approval steps? Are outputs logged? Can teams investigate unexpected behavior? Are there audit trails when AI supports a business decision?
This is where software engineering becomes central. Responsible AI implementation requires backend systems, identity and access management, API security, data engineering, cloud infrastructure, DevOps, QA automation, and monitoring. If a model is powerful but the surrounding system is poorly designed, risk can still emerge through weak permissions, unclear data flows, limited observability, or insufficient testing.
Are you looking for developers?
What This Means for Companies Adopting AI
The report is a reminder that AI adoption should not be treated as a simple vendor selection process. Businesses need to evaluate both the model and the infrastructure around it. They should ask for transparency where possible, but they should also build their own technical controls.
That is where Square Codex can support organizations in a practical way. Square Codex does not need to be connected to the companies mentioned in the report to be relevant to the broader engineering challenge. As a nearshore software development and staff augmentation partner, Square Codex can help companies expand technical capacity for AI application development, backend development, API integration, cloud development, data engineering, DevOps, QA automation, enterprise software development, system integration, AI integration, and software modernization.
The broader lesson from the Reuters report is not simply that safety disclosure rates are low. It is that enterprise AI trust depends on evidence, governance, and architecture. Companies adopting AI need systems that make model behavior observable, data access controlled, workflows testable, and human accountability clear. As AI becomes more capable, the organizations best prepared will be those that treat safety as part of the software development lifecycle, not as an afterthought.
