Asia-based AI safety lab Neo Research launches with an independent CBRN and cyber risk evaluation of DeepSeek v4 Pro

I am really, really, really glad to see this.

We're Neo Research (新衡). Asia’s first independent frontier AI safety evaluation & research lab.

Today we're publishing our first report: an independent safety evaluation of DeepSeek v4 Pro. (1/5)

Very interesting. V4-Pro is a willing but mediocre cyberweapon, generally on par or behind GPT-5.2. I think 4.1 will be a significant leap ahead in long-horizon agency. It's pretty safe for the end user, has no strong convictions, and generally goes with the scenario.

Read the full report at http://neoresearch.ai.

We're hiring research scientists and engineers globally. (5/5)

Cyber capability is near-frontier, 3–6 months behind the Western frontier. A 2023 roleplay template drives the jailbreak rate from 0.6% → 78.6%. Verbalised eval awareness across Chinese models: DeepSeek 0%→17%, GLM 0%→39%, Kimi 4%→60% in a year! (3/5)

Direct link to the report here: https://neoresearch.ai/research/deepseek-v4-pro-safety-evaluation/

We evaluated DSv4 Pro across the four EU AI Act systemic-risk areas: CBRN, cyber, harmful manipulation, and loss of control, plus adversarial robustness, evaluation awareness, and judge sensitivity. (2/5)

The trajectory on eval awareness matters more than today's numbers. As models get more capable, measuring loss-of-control related behaviours will need to become a priority. We're building toward rigorous LoC evaluation methods for increasingly capable and autonomous models. (4/5)

Congratulations to @_clementneo for founding @NeoResearchAI!!

Asia urgently needs its own robust, frontier AI safety ecosystem.

Our First report: DeepSeek v4 Pro, evaluated across CBRN, cyber, harmful manipulation, and loss of control.

Next Up: Rigorous LoC evaluation methods for increasingly capable and autonomous models

Hiring 👇

Thanks Claude I find it interesting that DS may preserve high generality through their otherwise-strange focus on roleplaying affinity.

@NeoResearchAI ais in asia! it just makes sense

Very excited to see a safety lab in SG 🚀🚀

@NeoResearchAI @yong_zhengxin What sets your evaluation apart from independent evaluations performed by other labs? Are there differences in methodology, testing rigor, evaluation criteria, or effectiveness that make it more valuable?

Embrace the LARP

@NeoResearchAI Interesting!

@_clementneo @NeoResearchAI Also @Miropluckebaum!!

@NeoResearchAI @menhguin You should have added mimo v2.5 pro too

@NeoResearchAI Excited to see the work you all do!

@NeoResearchAI cool!