AI Builders Digest

Vercel says its security agent patched 1,411 vulnerabilities in 24 hours

Guillermo Rauch says Vercel Security Reviews identified 6,817 vulnerabilities in 24 hours, including 580 rated critical or high, and an agent patched 1,411. He predicts AI will repeatedly rewrite human-made software to improve its speed, security, and correctness. Vercel Security Reviews are free through November 10.

• #agents
• #security
• #products

Replit expands its partnership with a major Japanese financial firm

Amjad Masad says Replit is powering core software products for Sumitomo Mitsui Finance & Leasing. In one example he shared, one person finished in a week a project that previously required six people working for six months. The partners plan to bring AI-powered software creation to more Japanese enterprises.

• #agents
• #enterprise
• #products

Kevin Roose calls for safety testing before AI models reach the public

Kevin Roose argues for an industry standard that tests models for risks such as autonomous cyberattacks before deployment, followed by a way to coordinate a slower pace at the frontier. “We should require AI companies to test their models before offering them to the public.” He also describes using a council of AI agents to review his book chapters; they caught factual errors missed by human editors.

• #policy
• #safety
• #agents

Swyx flags a model reported as faster and cheaper at the same eval score

Swyx points to a model reported as three times faster and five times cheaper while matching its FrontierCode 2 evaluation score. He calls it “distillation without tradeoffs” and expects more results like it from US-pretrained models.

• #models
• #evals

Cat Wu joins OpenAI to lead Core Products product management

Cat Wu says she is joining OpenAI to lead product management for Core Products. She says the team will build agents that support people and their teams while keeping users in control.

• #agents
• #products

Matt Turck sees a customer-service play in Dialpad and Ada’s combination

Matt Turck wants one AI assistant that can handle text or voice, sales or support, and take useful action at any hour. As an investor, he sees the combination of Dialpad and Ada as a major new entrant pursuing that vision.

• #agents
• #products

Peter Steinberger argues for agents that represent entire projects

Peter Steinberger questions whether individual limits still make sense as agents gain capability. He envisions an agent that represents a project, runs ongoing loops, and evolves its codebase, with people curating the automations.

• #agents
• #software

OpenAI has assembled a cyber team, Sottiaux says

Thibault Sottiaux says OpenAI has assembled a cyber team and expects to share more about its work in the coming weeks and months. He also points people interested in joining the effort to Clint Gibler.

• #security

Ryo Lu wants AI tools that leave room to decide what to make

Ryo Lu asks what happens when people can make anything but are less sure what they want. He argues that tools should help people form their own answers and make things together, rather than simply produce more output.

• #products