Tencent AI Chief Discusses Bioweapon Risks and Open Models

To manage these risks, leading labs employ targeted restrictions. For instance, Anthropic and OpenAI restrict their most powerful cybersecurity models to vetted partners via programs like Project Glasswing and Daybreak. Furthermore, updates to Claude's architecture introduced routing mechanisms where suspicious queries regarding cyber or biological topics are sent to weaker models. This aggressive filtering has occasionally caused friction; users reported legitimate tasks, such as analyzing MRI scans, being flagged as potential bioterrorism, prompting acknowledgments from Anthropic that their filters were sometimes too restrictive.

Conversely, open models offer distinct advantages during operational incidents. Yao pointed to a real-world cybersecurity event at Hugging Face where closed US models, including Anthropic's Claude, declined to analyze attack logs due to safety guardrails. As a result, the platform utilized the open Chinese model GLM-5.2 to parse the logs instead. While the actual attack was ultimately halted through traditional patching, server rebuilding, and credential rotation, the incident highlighted practical use cases for open architectures during forensic analysis.

Vulnerabilities and Regulatory Responses

A major challenge for open-source systems is the fragility of their safety filters. A technique known as abliteration allows individuals with access to model weights to remove trained refusal behaviors. Thousands of modified variants exist on public platforms, and startups even sell access to unfiltered versions. While current countermeasures make these modifications more difficult, they do not entirely prevent them. Experts suggest that a more robust strategy is withholding dangerous knowledge during the initial training phase.

Regulatory bodies are taking notice of these vulnerabilities. In China, the standards organization TC260 has acknowledged that open model safety mechanisms can be bypassed and is developing a dedicated security standard for open-source AI. Meanwhile, a draft export control framework covers exceptionally large models while exempting previously released open models. Neither the United States nor China has established a binding capability threshold that prohibits the publication of open models.

What it means for developers

Developers navigating this complex landscape need flexible access to a wide range of systems to handle different technical tasks, from routine analysis to complex reasoning. If you want to experiment with these technologies without managing multiple subscriptions, developers can try top AI models cheaply through one API at https://apixoai.online.

Meanwhile, major industry players continue to refine their release strategies. Tencent plans to launch its new model, Hy4, as an open-weight release alongside a comprehensive safety report. The company stated that Hy4's biological capabilities fall safely below high-risk thresholds, and that its hacking capabilities were actively filtered out of the training data using an approach designed to resist abliteration.

Source: Tencent's AI chief says AI bioweapons are coming, but closing off models isn't the answer — The Decoder. Written by the Apixo team from that report.