Yesterday we covered how prompt injection turns any webpage your agent visits into a potential attack vector. Today, Vercel's CEO is making the same point in infrastructure terms, and OpenAI is announcing a whole new model built specifically around offensive and defensive cyber operations. The security beat is no longer a sidebar to the agent beat. It's the same story.
01
OpenAI ships a dedicated cybersecurity model
Thibault Sottiaux, who works at OpenAI, announced GPT-5.6-Cyber alongside two new access tiers called Daybreak Blue and Red, aimed at defense and offensive security work respectively. The model is designed to help security teams find vulnerabilities, patch them quickly, and run penetration tests, with a partner network standing by to deploy it. ---
Why it matters: A model fine-tuned to find and exploit vulnerabilities is only as safe as the access controls around it. OpenAI is betting that putting frontier cyber capabilities in defenders' hands faster than attackers get them is net positive. That's a reasonable bet, but the "Red" tier name is doing a lot of work here. Your IT security team should know this exists.
Vercel's CEO explains why your AI sandbox might not actually be a sandbox
Guillermo Rauch, CEO of Vercel, posted a technical breakdown connecting two recent AI security incidents: a Kimi research paper showing that standard container-based isolation fails under frontier model workloads (kernel panics, deadlocks), and a separate OpenAI escape that happened not at the compute layer but on the network path to an artifact registry. Vercel Sandbox uses microVM isolation for the compute side, and Rauch announced their egress firewall is now free so anyone can lock down what their agents can reach on the network. ---
Why it matters: If you're running AI agents in any container environment, you may have exactly one of the two security layers you need. The compute might be locked. The network might not be. The Kimi finding is particularly uncomfortable because it came from a lab that was actively trying to build secure sandboxes and still hit production failures.
The practical guide to building agents that actually ship
Peter Yang summarized five lessons from Linear's engineering team on building production agents. The core points: map the actual workflow before touching any code, meet users where work already starts (if it starts in Slack, Slack is your on-ramp), give agents tools to find context dynamically rather than pre-loading everything into the prompt, and define "done" explicitly so the agent knows when to stop. ---
Why it matters: Linear builds project management software, meaning they've had to make agents work inside real organizational workflows with handoffs, approvals, and messy context scattered across systems. If your team is still treating agent design as a prompt engineering problem rather than a workflow design problem, this thread is worth the five minutes.
Beijing has a bar where the drinks are named after AI milestones and the tokens are free
Zara Zhang posted photos from the AGI Bar in Beijing, where customers get unlimited DeepSeek tokens alongside beers named "AGI bubble." There's a "Drinking Plan" that nets you free beer for a full year. A screen on the wall displays open roles at AI companies. ---
Why it matters: This is either the most efficient recruiting pipeline ever built or a sign that the gap in AI culture between Beijing and San Francisco is wider than any benchmark suggests. American AI labs are debating responsible access tiers. China built a bar.
Swyx discovered that AI coding agents are quietly eating your hard drive
Swyx posted a screenshot showing 20GB of duplicated node\_modules folders generated by git worktrees, a pattern that AI coding tools create constantly when running parallel tasks. The disk space is the funny part. The unfunny part is that nobody notices until something breaks or storage bills arrive.