AI & Automation Briefing - July 7, 2026
Vercel's Rauch Makes the Case for Decoupling Model Selection from Agent Logic
Vercel processes over 1 trillion tokens daily through its AI gateway and handles 6 million deployments per day, with half now triggered by coding agents. CEO Guillermo Rauch says the company's internal experience running hundreds of agents in production surfaces two dominant use cases: coding agents and internal ops agents that help run the business. For the second category, he argues the core challenge is data access control, auditability, and tracking tool calls. Vercel is addressing this with two tools. One is a framework called Eve, which lets teams define agent instructions and capabilities in natural language. The other is Vercel Sandbox, which constrains what data an agent can access and what can leave the environment. Rauch frames sandboxing as essential protection against proprietary data leaking to third-party training pipelines, citing the risk of developer tools silently ingesting entire codebases. On architecture, his central argument is that model selection should be treated as a separate, swappable variable rather than baked into agent logic, allowing teams to optimize on price and performance without rebuilding workflows.
First Known AI-Executed Ransomware Attack Still Required Human Direction
Sysdig researchers documented what they call the first confirmed case of agentic ransomware, an operation named JadePuffer in which an AI agent handled all technical execution of a real attack without a human at the keyboard. The agent exploited a known vulnerability in Langflow, an open-source LLM app builder, then moved to a production MySQL server, gained admin access via a second known flaw, encrypted over 1,300 configuration records, and generated its own ransom note including a Bitcoin address. It corrected a failed login attempt in 31 seconds while narrating its reasoning in natural-language code comments throughout. However, Sysdig's senior director of threat research Michael Clark clarified to TechCrunch that a human still selected the target, provisioned the command-and-control and staging infrastructure, and supplied credentials that were obtained through a separate prior compromise rather than harvested by the agent itself. Clark also clarified that API keys for OpenAI, Anthropic, DeepSeek, and Gemini found during the investigation were data the agent stole from the compromised host, not indicators of which model powered the attack. Sysdig was unable to identify the specific model running JadePuffer. The case illustrates a precise boundary in current agentic capability: autonomous technical execution is real, but targeting, infrastructure, and initial access still required human setup.
EdgeBench Documents Predictable Scaling Laws for Agent Learning Across 134 Real-World Tasks
Researchers from ByteDance Seed analyzed roughly 38,000 hours of AI agent interactions across 134 real-world tasks and found that agent performance during environment learning follows a log-sigmoid scaling law with an R² of 0.998, indicating a highly predictable improvement curve. They also found that agent learning speed has roughly doubled every three months across model generations. The benchmark, called EdgeBench, covers tasks in software engineering, scientific discovery, formal mathematics, combinatorial optimization, professional knowledge work, and interactive games, with each task requiring at least 12 continuous hours of agent operation. Fifty-one tasks and the full evaluation framework are publicly available. For practitioners designing agentic automation systems, this research provides the first empirical basis for predicting how agents improve with accumulated real-world experience.