We are excited to share that we supported the development of Tencent Hy4’s system card. It is the first standalone, safety-focused system card including publicly disclosed third party evaluations of an Asian model and marks a great first step towards harmonising international best practices on AI safety.
Replying to @TencentHunyuan
A special thanks to @aisafetysg and @NeoResearchAI for their valuable contributions to Hy4's safety evaluations, bringing additional perspectives on agentic safety, evaluation methodology, and emerging risks.
We appreciate the thoughtful collaboration and look forward to building on it!
We believe that shared methods and published results are how safety practice converges across labs and countries. To that, we provided perspectives on agentic safety, evaluation methodology, and emerging risks. We look forward to deeper and more extensive work with multiple frontier labs on testing and evaluations across model releases.
1
5
268
If securing frontier AI safety, developing international AI safety standards and working on frontier research interests you, we are hiring globally.
neoresearch.ai/careers/
5
127
We independently evaluated Kimi K3 across 7 misalignment categories and almost 30 benchmarks.
As models gain autonomy, we need to better understand how they use the access and authority they are given.
Three patterns stood out: overeagerness, metagaming and eval awareness. 🧵
These concerns extend across the frontier. As agents become more autonomous, we need stronger alignment, monitoring and access controls, as well as realistic tests of how agents behave.
Read the blog post and full report here:
neoresearch.ai/research/kimi…
1
2
40
Working on misalignment, monitoring or more realistic safety evaluations? We’d love to collaborate! Reach out to us at [email protected].
3
30
We present early safety results on GLM 5.2, currently the strongest open-weight model available.
Mostly it lands between Claude-family models and DeepSeek v4 Pro. But one number went down for the first time and we can't yet tell whether that's good news. 🧵

