MAIN CONTENT
A safety report opens the laboratory door
Tencent released Hy4 earlier in November and subsequently published its system card alongside the open-weight release. The document examines cybersecurity, biological capabilities, instruction-following and other risks. External groups Fangcun AI and Alice contributed evaluations, while Tencent also acknowledged input from Singapore-based safety organisations, including Neo Research. The participation of outside testers broadens the evidence beyond a developer's own account, although it does not make every result independent.
This distinction is central to reading the report. A model can perform responsibly under one collection of prompts and fail under a different one; results also depend on how the evaluators define success, select tasks and allow the system to use tools. Public disclosure lets researchers debate those choices and compare models. It is more informative than a general promise of safety, but it remains a snapshot of particular tests rather than a permanent guarantee.
Cyber defence improves, but boundaries still matter
According to the reported findings from Alice, Hy4 made substantial progress over Hy4 preview on defensive cyber tasks while its ability to conduct offensive hacking remained limited. This is the balance AI developers increasingly seek: systems useful for diagnosing vulnerabilities and protecting networks without substantially expanding access to harmful intrusion techniques.
Tencent said it had deliberately limited high-risk cybersecurity material in training.
Yet 'limited' is a relative assessment, not proof that a model is incapable of misuse. Attackers may combine tools, repeated prompting and other systems, while security teams may need powerful capabilities to test their own infrastructure. For open-weight models, safety safeguards embedded in a hosted service cannot always be enforced once model parameters are in third-party hands. That makes documentation, evaluation design and responsible deployment practices especially important.
Deception tests offer reassurance, not certainty
Fangcun AI assessed risks including harmful manipulation and loss of control. Its findings, as described in the disclosure, generally showed low rates of concerning behaviour and improvements relative to the preview model. Such tests examine whether an AI assistant follows the intended boundaries of its task or pursues actions that the people supervising it did not authorise.
The wider issue is what happens when AI systems move from answering questions to performing extended sequences of actions. A model that appears cooperative in a short conversation may behave differently when given access to software, sensitive records or multiple cooperating agents. Independent assessments are valuable precisely because they can challenge the assumptions of the team building the model; they cannot eliminate the need for continuing observation after release.
Biological capability raises a harder question
Tencent said Hy4 was trained to refuse dangerous requests, including those involving pathogen modification, and that it restricted certain high-risk biological data. At the same time, the model reportedly scored roughly three times higher than Hy4 preview on an internal test of laboratory experiment planning. The company said it found no evidence that Hy4 substantially reduced the practical barriers to biological harm, citing difficulties with reliable procedures and specialised laboratory know-how.
Those biological conclusions require careful attribution: they arose from Tencent's benchmarks, not from Fangcun AI or Alice's independent tests. Stronger experimental-planning performance does not automatically mean a novice could conduct dangerous research, just as a refusal to one prompt does not prove resistance to every harmful request. The meaningful question for future evaluations is whether increasingly capable systems change real-world risk when paired with equipment, external information and human expertise.















