Tencent filters dangerous training data for Hy4 model, rejecting open-closed safety tradeoff
Source published ·Modelwire updated
Original coverage: The Decoder ↗·How Modelwire adds context

The development
Tencent's chief AI scientist has articulated a strategic position on model safety that rejects both extremes: fully open systems vulnerable to adversarial modification and heavily restricted models that over-filter legitimate use. Rather than relying on post-hoc safety layers, Tencent is experimenting with upstream data curation for its Hy4 model, removing dangerous knowledge during training itself. This approach signals a shift in how major labs think about containment, moving beyond binary open/closed debates toward architectural choices that embed safety into model development. The framing around bioweapon risk underscores rising concern among insiders about dual-use capabilities in frontier systems.
Modelwire’s AI-generated summary of coverage from The Decoder.
Modelwire analysis
Analyst takeOur AI-generated reading of the wider context and the next developments to watch.
Yao's framing sidesteps the real constraint: upstream data curation only works if you control the training corpus, a luxury unavailable to labs racing against open-weight competitors. The bioweapon rhetoric may be real risk, but it also conveniently justifies architectural choices that lock in Tencent's competitive position.
This announcement arrives as Chinese open models (DeepSeek, GLM) have already demonstrated parity with proprietary systems on offensive capabilities (per November 7th coverage), collapsing the moat that once justified closed architectures on safety grounds. Tencent's pivot to 'embedded safety' mirrors Anthropic and OpenAI's recent acceptance of deployment risk (early November), but with a Chinese twist: framing containment as a technical problem solvable through training rather than access control. The tension is sharp: if open models already match closed ones on dangerous capabilities, Tencent's data-curation approach either works (and should be open-sourced to matter) or it doesn't (and the announcement is positioning, not innovation).
If Hy4's next version demonstrates measurably lower bioweapon-relevant capability than comparable open models on identical benchmarks, the upstream curation thesis holds. If capability parity persists despite the approach, Tencent has chosen a safety narrative that doesn't scale, and the real differentiator remains cost and inference speed (where Chinese labs already lead).
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·The Decoder
Open models close the gap on proprietary AI cyber attack capabilities
The competitive moat between proprietary and open-source AI models in offensive cybersecurity is collapsing. GPT-6 Sol Daybreak achieves 80.8 percent success on simulated network attacks, but DeepSeek V4.1 Flash reaches 72.8 percent at one-tenth the cost, eliminating the performance justification for closed systems. This convergence matters because open weights cannot be recalled or restricted once…
MentionsTencent · Vinces Yao Shunyu · Hy4 · Anthropic · Claude
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. The Decoder originally reported this story as “Tencent's AI chief says AI bioweapons are coming, but closing off models isn't the answer”. The full content lives on the-decoder.com. If you’re a publisher and want a different summarization policy for your work, see our takedown page.