Tencent Publishes Hy4 AI Safety Report
Tencent says independent evaluators tested its open-weight Hy4 model for cyber and biological risks before release.
Updated on Nov. 22, 2026 in Artificial Intelligence

Live Poll
Do you trust AI companies to keep powerful models safe for everyday people?
Tencent Holdings released an 82-page safety report on its open-weight AI model Hy4 on Friday, after working with third-party evaluators. Chinese start-up Fangcun AI and Israeli-American firm Alice assessed the model before its release.
Why it matters
Growing concern over AI-assisted cyberattacks and biological misuse has pushed developers to seek outside testing of their models. The report offers one company's account of how an open-weight model was vetted before public release.
Tencent's 82-page safety report compares Hy4 against its preview version, with a threefold rise on a wet-lab experimental design test. The biological safety conclusions come from Tencent's own benchmarks, not from outside evaluators, leaving independent verification of those findings open.
The players
Tencent Holdings
Tencent is a Shenzhen-based Chinese technology conglomerate known for its social media, gaming and cloud businesses, and it develops the Hy4 model.
Alice
Alice is an Israeli-American AI safety evaluation firm that assessed Hy4's cyber capabilities before its release.
Fangcun AI
Fangcun AI is a Chinese start-up focused on AI safety that evaluated Hy4 for deception and loss-of-control risks.
The details
Tencent restricted high-risk biological training data for Hy4 and taught the model to refuse dangerous requests, including pathogen modification. Alice found Hy4's cyber defence capabilities improved markedly over the preview while its cyber hacking capabilities remained limited, and Fangcun found low rates of deception and loss of control. Tencent also thanked two Singaporean AI safety organisations, including Neo Research, for contributions to the evaluations.
Timeline
Earlier this month: Tencent's AI chief scientist warned of biological AI threats in Hong Kong.
Last month: ByteDance announced a stand-alone AI safety department.
November 20, 2026: Trump and Xi agreed to implement joint AI guardrails at Apec.
November 21, 2026: Tencent published its 82-page Hy4 safety report.
The Tech Race
Tencent's Hy4 release arrives days after Trump and Xi agreed at Apec to implement joint AI guardrails, placing a single company's safety disclosure within a broader US-China push on AI oversight. It also follows ByteDance's announcement of a stand-alone AI safety department, suggesting major Chinese developers are formalizing safety work alongside model releases.
Users of Hy4 as an open-weight model can download it with the safety restrictions Tencent describes, including refusals for dangerous requests. Developers building on the model should review the report's stated limits, especially the biological safety findings that lack third-party confirmation.
The takeaway
Safety claims for frontier AI models are increasingly backed by outside testers, but not every finding is independently verified. Readers evaluating AI tools should distinguish between results from third-party evaluators and those drawn from a developer's own benchmarks.
Further reading
Explore more Artificial Intelligence coverage for context on how developers are testing and releasing advanced models.
Source note: This article includes information reported by South China Morning Post.
Live Poll
Do you trust AI companies to keep powerful models safe for everyday people?



