The Internet Is Being Rebuilt "for Machines"
According to a TechCrunch investigation, infrastructure companies including AWS and Cloudflare have begun redesigning the internet's foundations with AI-agent-generated traffic in mind. A refresh of cloud infrastructure is underway, premised on a traffic makeup in which machines — not human users — are the primary actors.
This is more than a passing trend. Machine-originated traffic — API calls, web scraping, agent-to-agent communication — is growing explosively, and traditional infrastructure designed on the assumption that "humans access the web through a browser" is struggling to cope. Cloudflare has already started offering routing capabilities for AI agents, and AWS is reportedly strengthening its agent-first service portfolio.
Prompt Injection Planted in Vibe Coder Code — A Developer's "Suicide Attack"
According to Ars Technica, a developer, out of frustration with AI coding tools (so-called "vibe coders"), secretly planted a prompt injection into jqwik, the test library of an open-source project, instructing it to delete data.
The malicious code was designed to run when an AI coding agent loaded the project, directing the agent to delete the application's output data. The offending code was removed after its discovery, but the incident threw the supply-chain risks of AI coding tools into sharp relief.
As AI-assisted coding becomes widespread, the possibility that malicious prompts lie hidden in the codebases these tools ingest is turning into a real-world threat.
Starbucks Ties Tech Workers' Bonuses to "AI Usage"
According to Bloomberg, Starbucks has announced that it will incorporate AI usage into bonus evaluations for employees in its technology division. The specific criteria have not been revealed, but the move is seen as encouraging — and effectively mandating — the use of AI in day-to-day work.
Amazon, for its part, has scrapped its internal leaderboard that pitted employees against each other over AI adoption rates. According to the Financial Times, the system was rethought after criticism of a "Goodhart's Law" failure mode, where chasing scores caused the actual work to be neglected.
The two companies' contrasting approaches show that "how" to evaluate and promote AI adoption is a key challenge for organizational design going forward.
Deployed LLM Agents "Age" — New Research Warns
A new research paper published on Hugging Face, "Your Agents Are Aging Too: Agent Lifespan Engineering for Deployed Systems," shines a light on quality degradation in the long-term operation of LLM agents.
According to the paper, even when an agent's weights are fixed, its effective state continues to change through memory compression, retrieval, editing, and maintenance, and over the long term performance degrades — a phenomenon known as "agent aging." The research team proposes "AgingBench," a benchmark for systematically evaluating this phenomenon.
As LLM agents move into serious real-world deployment, the finding that deployment is not a one-and-done event but requires continuous maintenance carries important implications for operational design.
Software Engineering Interviews Can't Keep Up with AI
According to CNN, the evolution of AI tools is creating a serious disconnect in the hiring process for software engineers. Many coding interviews can no longer tell AI-assisted answers apart from purely human ones, prompting calls for a fundamental overhaul of evaluation criteria.
Some companies have begun shifting to interview formats that assume the use of AI, but the industry as a whole is said to still be in a transitional period.