Our findings show that our fine-tuned BrowseSafe model outperforms both off‑the‑shelf safety classifiers and frontier LLMs used as detectors. These gains are possible through fine-tuning on BrowseSafe-Bench data, allowing us to bypass the reasoning latency of larger models.
RESEARCH
-

DeepSeek-V3.2 Next-Generation Foundation Model Launches on Poe
By
–
Now on Poe: DeepSeek‑V3.2 This next-generation foundation model, powered by @novita_labs on Poe, is designed to unify high computational efficiency with state-of-the-art reasoning and agentic performance. (1/2)
-
Terminal-Based Agent Control and Harbor Evaluation Framework
By
–
Key takeaways:
• terminals > GUIs for stable agent control
• task design inspired by SWE-bench, but with a more general abstraction
• eval + RL need the same “rollout” substrate, so they created Harbor
• Harbor = a unified framework for scalable parallel deployment
• TB2 is -

Terminal-Bench Team Interview Highlights Recent Chat
By
–
ICYMI: Our chat with the Terminal-Bench team dropped recently.
-
MCP Security Paper: Risks and Controls for Enterprise Adoption
By
–
We just released a paper on Model Context Protocol security – joint work with Herman Errico (
@TrustVanta ) and Shanita Sojan (
@Darktrace ). As MCPs get adopted, companies need to understand the risks + controls needed in order to be secure and stay compliant. In this paper, we -
Google’s Legacy and Generative AI’s New Challenges Today
By
–
In 1998, Sergey Brin and Larry Page, were still students at Stanford when they intro'd Google search prototype in their now-famous paper. Now, decades later, generative AI is creating new challenges and reminding everyone of old ones.
-

NeurIPS Happy Hour: Neural Operators and Scientific AI Research
By
–
Join our happy hour meetup today at @NeurIPSConf to chat about AI+Science and AI+Math with me and my team from @Caltech at: Achilles Coffee Roasters Gaslamp, San Diego. 4:45pm – 6:30pm We have a number of papers at NeurIPS on Neural Operators, Inverse Problems, Scientific
-

Mistral 3 now available under Apache 2.0 license
By
–

BREAKING: Mistral AI has released Mistral 3 Large and Ministral 3 in various sizes under the Apache 2.0 License! – Ministral 3 3B-25-12
– Ministral 3 8B-25-12
– Ministral 3 14B-25-12
– Mistral Large 3 (675B) -

CuspAI expands to London Kings Cross and welcomes new Chief Scientific Officer
By
–
We’re also going to be expanding our UK footprint with a space in London’s Kings Cross from January 2026. Our London base is the latest as CuspAI continues to scale, and is part of Cusp’s growing global footprint including Cambridge, Amsterdam and Berlin. Max Welling (@wellingmax) Our company @cusp_ai is incredibly excited to welcome globally renowned researcher in materials science Professor Aaron Walsh as our Chief Scientific Officer. — https://nitter.net/wellingmax/status/1995854008494293262#m
→ View original post on X — @wellingmax, 2025-12-02 14:01 UTC
-

CuspAI Appoints Professor Aaron Walsh as Chief Scientific Officer
By
–
Our company @cusp_ai is incredibly excited to welcome globally renowned researcher in materials science Professor Aaron Walsh as our Chief Scientific Officer.
→ View original post on X — @wellingmax, 2025-12-02 13:54 UTC
