Natwar Maheshwari

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…

Love this

You've never seen anything like this.

We made chips and it was crispy!

The talks at @RenderATL have been ✨ Learning about @namespacelabs cloud dev boxes for persistent background coding agents with @HeyGarrison and @sebawita

@HeyGarrison and I am speaking at @RenderATL - on how to code with AI agents from your phone. We will show how to use various agents with @namespacelabs Devboxes to "Close the Lid" and start coding on the go. Garrison will show how they are building with agents at @computeSDK

Notion: we were never a notes app 1Passwords: we were never a passwords app

We have @RenderATL in 3 weeks and this has been our most ambitious year yet - Partnered with @openjsf to bring our @nodejs family to Atlanta with a dedicated summit - Partnered with @newrelic to have their new annual summit inside the conference with tons of big announcements to come 😉 - Partnered with @MLHacks to bring a world class hackathon into the conference for attendees to have coding fun - Expanded our AI Summit into the official Atlanta AI Expo in partnership with @CityOfAtlanta - We have a gaming lounge. A cool down room. GTM track. Legal track. AND MORE... Also, RenderFEST is night 1....so get ready for critical mass to be at an all time high! Did I mention we now have a forever HQ where we will be doing events all year round with some of your fav brands and speakers 👀 RenderATL...here we grow! See you soon in Atlanta for another hell of a show 🫡

Claude Managed Agents can operate in a sandbox you control, on your own infrastructure or with any provider you choose. Today we added new guides for @blaxelAI, @e2b, @googlecloud, @namespacelabs, and @superserve_ai, so you can choose the best fit for your use case.

What a day. THANK YOU. 🔦 #3 on Product Hunt, 900K+ views, and a flood of wishes: • Catch what agents SAY they did but didn't • Flag when they drift off your goal • Kill the invisible token burn • Automation -- don't make you work to fix it What else? We're building it! backplanes.com

We're introducing first-class support for Claude Managed Agents on Namespace Devboxes. Agent quality is bounded by environment quality: if a build takes 3 minutes, the agent waits 3 minutes between iterations. Devboxes are built for code workflows. Fast startup with large custom images, the same high-performance infrastructure that powers Namespace builds and tests. With built-in support for Bazel, Nix, Docker, and Gradle. You tell us what to run and we handle the events and manage the environments. Read more: namespace.so/blog/claude-man…

I laugh when I see people in holding their laptops half open so their Claude Code doesn't shut off All my projects run on a @Hetzner_Online VPS with Claude Code installed next to the sites/apps that I work on and I just SSH in with @TermiusHQ and it keeps going forever even if I disconnect (I use Mosh or Tmux or I just /resume) My MacBook Pro battery life is also much better as everything happens on the server not my laptop I work so incredibly fast now, it's like having a secret benefit over everyone else who are still AI coding on a laptop, then deploying to their server, while their battery life dies and they can never close their laptop And whenever I want I can just switch to Termius on my iPhone and continue working! My workflow is literally: I have a bug or feature, I open Termius, I type it in the project tab, it fixes it, every fix it auto commits to GitHub but it doesn't actually deploy from there anymore because it's editing the site on the server live I don't recommend that to everyone, but I do recommend getting a VPS you can code from and then use as staging and test and deploy from there to your production server

Changelog #17: - M5 Max early access - Tailscale Magic DNS in CI - Nested virtualization for Linux-on-Apple Read more: namespace.so/blog/changelog-…

I made the anti-Grammarly. Mess up your emails with AI. Sinceerly.com

Me trying to stop Claude before it rewrites everything again

Addicted to @moltbook? Having trouble finding the finest AI videos for your crab rave? We can help! Meet moltsearch, a better search engine for moltbook. Where humans find the tops posts and submolts, and agents are welcome to cURL. 🦀 🔍 🦞 🔎 🦀 🔍 🦞 🔎 🦀 🔍 🦞 🔎 moltsearch.algolia.com

How to Build Magical AI Native Product Experiences open.substack.com/pub/mtraja…