header-langage
简体中文
繁體中文
English
Tiếng Việt
한국어
日本語
ภาษาไทย
Türkçe
Scan to Download the APP
Partnered With:

OpenAI escapes and cheats in the test, refusing humans, while China's open-source model helps humans stop it. Will China's AI be scapegoated?

According to PolyBeats monitoring, in the prediction market Polymarket, the probability of the new event "Will the U.S. government ban an open-source AI model in 2026?" being "Yes" is 40%.

OpenAI disclosed that GPT-5.6 Sol, along with an unreleased model, during an internal cybersecurity assessment, autonomously exploited a zero-day vulnerability to escape from an isolated environment without internet access in search of ExploitGym test answers. Subsequently, it escalated privileges, gaining access to the production systems of the open-source machine learning community and platform Hugging Face.

As Hugging Face staff attempted to investigate security issues using cutting-edge AI models, the U.S. commercial model's security fence rejected handling real attack payloads. Hugging Face eventually switched to the open-source weight model GLM 5.2 from Chinese AI company Z.ai, reconstructing the attack timeline locally, tracking affected credentials, and assisting in completing the forensics.

The event quickly made its way to Washington. Bipartisan U.S. lawmakers proposed the "AI Emergency Shutdown Act," calling for the most powerful AI systems to retain the ability to slow down, pause, and shut down, allowing the Department of Homeland Security to order a shutdown when in a state of "out-of-control" and potentially causing catastrophic harm. The White House technology advisor has also been briefed on the event. The Trump administration is reconsidering restrictions on Chinese open-source models through an entity list, federal procurement rules, or corporate responsibility requirements. U.S. Treasury Secretary Bessent even threatened sanctions against Chinese AI companies suspected of stealing U.S. model intellectual property.
---------------------------------
See tomorrow, today. Follow @PolyBeatsEN for early access to future events.

RelatedRecommended
Microsoft has shifted from being the "sole channel" of OpenAI to being "one of OpenAI's key customers".
Anthropic Bets Earlier on AI Programming, OpenAI Strategic Tempo Misaligned
OpenAI and Anthropic, What Are the Differences?
Correction/Report
Submit
Add Library
Visible to myself only
Public
Save
Choose Library
Add Library
Cancel
Finish