🚀 DeepSeek-V4.1-Flash-1017 is here! Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API. 🛠️ Optimized for @DeepSeekHarness, Tencent CodeBuddy, WorkBuddy, and more. Upgrade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model. 1/4

Muse after I give it access to my credit card and the open internet

Every serious frontier tech company - big or small - should assume it has insider threats. And they don’t need to be that sophisticated. I saw it in government. I saw it with crypto. OpenAI is the latest. Founders tend to assume good intentions while speaking about some of their most sensitive information out loud: capabilities, unreleased products, strategic plans, government relationships. Employees are the most obvious vector. The more sophisticated operations target founders, execs, and vendors through romantic, family, and social relationships. And on small teams, one compromised person can do a lot of damage. With former federal prosecutors and ex-FBI agents as part of our team, we spend a lot of time on these issues. Here’s some of what we’re telling our founders: 1. Expect to see more private-sector versions of SCIFs and clearances, where the most sensitive conversations and work are physically isolated and under stricter access controls. 2. Agencies may know about a specific insider threat at a company but can't share that info unless someone at the company holds a clearance like TS/SCI. Lead times to get a clearance are long. Plan now. 3. Restrict access to sensitive meeting schedules, disable unused ports on conference-room equipment, and regularly sweep rooms used for confidential discussions for surveillance devices. You don't need OpenAI's budget to be ready. But you do need to start.

I have just started to contribute to Quest, @electriccapital's open-source agent harness, where we just integrated @near_ai's trust infrastructure! Quest is a self-hosted harness where you plug in your own inference providers, connect the services you already use, and let agents work on them behind approval gates and cost reporting. When you send a prompt to a hosted model, you are trusting whoever runs the box: they can read the prompt, read the answer, and quietly serve a different model than the one you asked for. For an agent that is reading your emails, your docs, your internal APIs, that is quite a lot of trust with your most confidential data. @near_ai cloud runs part of its catalog inside Intel TDX and NVIDIA confidential-GPU enclaves, where the hardware attests to what is running and NEAR says even its own operators cannot look inside. If you or your organization deals with confidential data, legal documents, company finances, and client records, Quest is definitely worth spinning up, so that your data stays with you.