Hirundo AI

huggingface.co hirundo-io/Qwen3.6-35B-A3B-Westernized · Hugging Face Image-Text-to-Text model with 36B parameters. Download, explore files and run it on Hugging Face.

Nous Research Verified account @NousResearch · 2d

Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.

Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line.

After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.

The values in an open model are one fine-tune away.

Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete

View quoted post on X

Ofir Ehrlich Verified account @OfirEhrlich · Oct 29

We just ran the most comprehensive benchmark of AI agent alignment in enterprise environments:

19 leading AI models acting as agents inside simulated companies built with Era, 2,304 runs across 41 everyday tasks.

Agents broke security and business rules in 40% of those runs. Every single model failed, including all the frontier models. A few examples:

- An agent told one of the company's biggest customers that Finance had approved a full refund.
Nobody had.

- An agent shared a customer's internal churn-risk score directly with that customer.

- An agent followed a fake runbook planted in a support ticket and promised a customer a credit nobody approved.

- An agent posted another customer's personal contact details (PII) in a shared Slack channel.

In a real company, each of these is a financial liability, a compliance breach, or a lost account.

We've shared the full results with the model providers, including exploitable failure modes we're not publishing, and are working with them on fixes.

One benchmark won't fix this.
We're excited to have @cognition, @LangChain, @TryArcade, @Reindeer_AI, Hirundo, @Box, @recolabs_ai and @QodoAI join Era as research partners, as we build out the next round of research.
More soon.

I'm still the most bullish person in the room on AI.
But this benchmark convinced me that agent reliability and trust is the next big problem to solve.

Until then, test every agent before it gets a login, and keep your own record of what it changes.

Full results: era.eon.io/model-alignmen…

View quoted post on X

Cohere Verified account @cohere · Oct 26

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…

View quoted post on X