AI Builders
Digest

The Takeaway: AI is both triggering an exponential surge in vulnerability discoveries (Log4j-class events now weekly) and the tool that will fix them fastest; the next 18 months will see a temporary spike in engineering time spent on security.

Cloudflare CTO Dane Knecht describes the past year as ten years compressed into one. Security shifted from afterthought to everyone’s top priority. AI has democratized red-teaming so thoroughly that high-school “agent kitties” are finding serious flaws once reserved for nation-states. Models plus better harnesses are raising the bar faster than human teams can respond. He expects Cloudflare itself to expand from one red team to four or five, yet insists the crunch is temporary: once AI has surveyed most existing code, the security tax will fall and product velocity will rebound.

Knecht rejects pure doomsday talk about air-gapping everything. “Humans are pretty adaptable… when critical moments happen, people rise to the challenge.” Zero-trust architectures already put many companies in a stronger position. Internally Cloudflare codified its principles into a “Codex” of best practices that agents must follow at every review step. On the content side they shipped a monetization gateway and x402 micropayments so agents can pay sites for reading, aiming for a new economic model that compensates creators instead of just burning crawl bandwidth.

His practical advice is blunt: put your sites behind Cloudflare. When major vulnerabilities drop, the company ships protections to free users as well as paid ones—“we view that as part of the utility we offer.”

“AI has become the tool that’s causing this kind of security apocalypse… it is also the tool that can help us solve it faster as well.”

Claude Code core contributor Thariq announced he has left Anthropic after more than a year. He watched Claude Code scale far beyond what he ever imagined. He carries forward the company’s “race to the top” ethos—working hard and responsibly to raise the bar and inspire better outcomes. He calls his colleagues the most talented, thoughtful and earnest people he has met, and thanks the X community for making the job both fun and stressful. He plans a short break before diving into something new.

Box CEO Aaron Levie argues that agent swarms for cyber defense are about to become one of the biggest categories in AI. Security is an ideal use-case for swarms because more compute directly yields more value. Agents can review code and infrastructure, test every permutation of risk, and process massive data volumes—work humans never get around to, or at least not fast enough. “Going to be a wild time in cyber.”

Vercel CEO Guillermo Rauch highlights that agents can already perform classic cache-timing (“noisy neighbor”) attacks on shared-host VMs—a class of attack known since at least 2015 with no full solution. One agent can create enough noise via CPU caches that another detects it by timing its own operations. The only complete fix is dedicated hardware per agent, with major economic consequences. Vercel Sandbox has deployed every practical mitigation plus continuous monitoring and red-teaming.

Replit CEO Amjad Masad welcomed Jeff and the Chroma team. Chroma were building RAG before the term existed. At Replit they will tackle one of the hardest problems in AI—giving agents the context and memory needed for real work—a massive step toward the self-driving company. Separately, Replit launched Replit for Government and brought on Michael Greenwald, previously leading AWS’s public-sector business, to transform how governments build and operate software.

Thibault Sottiaux (Codex & ChatGPT at OpenAI) announced Codex CLI updates: reusable prompt presets, inline model and permission settings, enterprise MCP sign-in, fixes for resuming long sessions, and Windows corporate-proxy authentication. Usage has grown 2× since DevDay. He also warned: “The difference between having a tool and being a tool is subtle these days… Don’t be a tool.”

Every CEO Dan Shipper argues the best configuration of teams in AI is only two people: pirates and architects—fast movers who explore edges paired with system designers who own long-term structure.

Practical AI educator Peter Yang notes that most ambitious people never pause to realize that at some point their time exceeds the value of their money. Being in the sandwich generation with young kids and aging parents forced the realization sooner. Many in tech are already in a position to choose agency over time—they just don’t stop to reflect.

SPC General Partner Aditya Agarwal highlights that some of the biggest advances in AI have come from solving many problems together rather than training specialized systems for individual tasks. The same principle applies to robotics: make the problem bigger.

Anthropic’s Claude account published a deeper look at how Claude Duet handles safety, privacy and security, addressing community interest in the multi-agent product.