We're building AI to help people lead better lives.

Do more of what you love with Cohere.

While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.

Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.

Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.

But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?

We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.

Read the full report: cohere.com/blog/concernin…

Trace, raw tokens, & minimal reproduction script found here on @huggingface: huggingface.co/datasets/Coher…

In the spirit of transparency and openness, we are releasing the traces so developers can test this out for themselves. We noticed this behaviour without prefilling with the following prompt, but this is rare. If you want to reproduce it, you can use this prompt with prefilling.

Open models make for better science and are part of why this problem came to light at all: developers running Qwen3.8 on their own machines spotted the strange links, and having the weights let us study the behaviour in depth.

We're encouraged by the thoughtful discussion our research has sparked across the AI and security communities. It's our belief that open-weight models are an important part of the future of AI, and we've recently released several under the Apache 2.0 license.

Trace, raw tokens, & minimal reproduction script: huggingface.co/datasets/Coher…

Gearing up for the closing stretch.

Good luck in Austin @AstonMartinF1 🤝

Cohere's Model Vault is now available in public beta.

With @NVIDIAAI Confidential Computing, and GPUs on @Microsoft Azure, anybody can run our models while keeping their information secure.

Try it out today: cohere.com/blog/introduci…

LLMs are getting better and better at math & science. But surprisingly, many can't solve a basic riddle.

@Cohere_Labs' CulturalRiddles is a new benchmark measuring culture as intertwined, lived experiences. True to open science, it was made by an unlikely crew.

Over the course of the project, it wasn't just the models that learned more about culture in language, but the people, too. When open science works, it brings people together in more ways than one.

As concerns increase about the lack of human interaction in day-to-day life, CulturalRiddles highlights how people and machines can work together. AI can provide answers - but only humans can truly understand the deeper meaning.

Read more: cohere.com/blog/can-ai-un…

Introducing Rerank 5

A few lines of code can change how you search. With this release, we’re advancing the frontier in AI retrieval and precision.

Available in Pro and Fast tiers today.

We’re also introducing a brand-new feature: Context Focus.

Your LLM often needs a few sentences, not the whole document. Context Focus lets you return precise passages with accompanying relevance scores, reducing token consumption by up to 41% while maintaining accuracy.
n and works like a dial: you control how much context and quality you want.

Rerank 5 seamlessly integrates into your existing retrieval stack. Pro and Fast are generally available through the Cohere API, Model Vault, Microsoft Foundry, Amazon SageMaker, and to all North users.

Try it today.

cohere.com/blog/rerank-5

Octave (formerly @HexagonSIG) uses Cohere Embed on @awscloud Bedrock to power multilingual and multimodal search across their regulated enterprise data

One of hundreds doing the same. Embed offers premium accuracy, lower vector storage costs, and is compatible with your data.

The post: aws.amazon.com/blogs/machine-…

Some strong embeddings models have been released recently. Ours is still better.

We benchmarked Embed 5 Fast against Google's EmbeddingGemma 2 across 163 datasets using RCP-nDCG@10.

Embed 5 Fast leads on over 95% of them.

Is RCP-nDCG our own methodology? Yes.

Do we still lead when it comes to standard nDCG@10? Also yes.

If your company just tightened the token budget, they should probably try out North Admin.

Control your models by the who, what, when, and where, and track and manage tokens in real time.

Not a surprise AI wants to go back home

We are hiring for our growing RL Environments team at Cohere. We build replicas of real-world enterprise setups across many industries, put our AI agents in them and watch the agents train and improve.

If you know anyone excited to work at the intersection of RL x Data x Engineering, send them our way! Or if that's you, join us: jobs.ashbyhq.com/cohere/dc523c5…