Jake

taking control of a sandbox in an AI product is such a poor experience. every click has visible latency. this screenshot is from @DevinAI. same experience from @daytonaio in @puffleai. our team has started testing UI locally because of this, which is really annoying. have any sandbox people solved for this?

"software is over" nope. think bigger! aws minus hardware is what? global shipping runs on what? public sector paperwork? even a "just" a medium size bank? we only had software for the smallest of all problems that could be solved with software, because we were severely limited by how inefficient it was to make. the scale ceiling was basically your phone's OS, or a browser, or a game. bigger things like automating a city or a medium-sized organisation's business we couldn't even attempt. and now things that used to take large teams many months can be done by one engineer in days. this is the BEGINNING of software! just imagine the possibilities! an OS for supply chain or a bank that runs itself any organisation of 1000+ people that _actually works_ it's been the holy grail since the dawn of computing but was beyond three level of capabilities of even the most sophisticated software companies guess what, now _that_ is going to be software and what we used to call software becomes ready made low-level components that just work, bundled into libraries or coming as part of bigger packages

⚠️ Running an AI harness outside of a sandbox? ⚠️ No tight least-privilege access spec'd for that sandbox's workload? ⚠️ no logs of net access & env probing for audit & privilege reconciliation? ⚠️ no strong sandbox sec bndary (not a VM)? ⚠️ real creds in the sandbox? 😟😟😟

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/concerning-b…

The list of patched software versions continues to grow

2,500 of you will have a Rex beta invite in your inbox in the next ten minutes