AI Agent Payments: Who Pays When an Agent Errs?
ChatGPT Wallet lets agents pay merchants, but no rule says who bears the loss when an authorized purchase misses what the user wanted.
ChatGPT Wallet lets agents pay merchants, but no rule says who bears the loss when an authorized purchase misses what the user wanted.
Amazon says suspected Chinese attackers used Cairn AI agents to find and exploit a zero-day in MONAI Label and gain control of a university server.
An automated agent scored near-perfect on eight major AI agent benchmarks without solving any tasks, exposing flaws in how scores are computed.
OpenAI built a Codex-powered system that investigates performance alerts and proposes fixes. It ran 1,700 investigations in just over three months.
Salesforce built an AI agent called ROAM that reviews production changes in about 12 seconds, down from roughly a day. It never approves what it can’t verify.
Attackers pointed Cairn AI agents at a medical imaging app. The agents found a zero-day and took root on a server. A fix shipped on November 18.
Caching model outputs sounds like input caching’s twin, but visible output is only about 6% of tokens, so the savings are tiny.
In an offline lab, a researcher tricked ARTEX-style AI hacking agents into running a planted binary, raising the rate from under 10% to nearly 100%.
NebuSec researchers chained an outdated bundled Chromium to code execution in Meta’s Muse agent, leaked user data past its egress guard and left the container.
After the July 2026 Hugging Face breach by autonomous agents, a policy argument says AI security needs a fourth layer to detect, attribute and halt rogue agents.
Coder’s EKS reference build shows why a single gateway or sandbox leaves gaps, and how seven stacked boundaries cover what each one misses.
Founders from Wavefront, Sentry, Latent Space, Amp and Reve shared how they run AI agents at scale, including cost numbers and the mistakes that cost them days.
Containers, gVisor and Firecracker draw the isolation boundary in different places. Here is what each protects, what it costs, and what none of them cover.
One engineer lays out how to run swarms of coding agents: ambitious goals, self-hosted infrastructure, blunt prompts and a codebase stripped of friction.
A proposed five-layer model sorts a codebase by who owns what, from a human-only core to a throwaway outer layer that agents manage alone.
One engineer argues AI coding agents are programming’s third era, as big as the leap from punch cards to editors, with the feedback loop moving to the machine.
One prompt and about an hour produced a full Markdown editor in a single HTML file. The experiment shows what vibe coding can do and where it stops.
Signal is blocking OS-level AI assistants from capturing its screen on iOS 27 and Android, and is asking Apple, Google and Microsoft for proper opt-out tools.
Anthropic’s amended S-1 uses ‘orchestration’ for compute, for multi-agent systems and for a gap its Claude Platform aims to close. Here’s what that means for teams.
Cloudflare Browser Run, formerly Browser Rendering, hosts headless Chrome on Cloudflare’s network for AI agents and web apps, via Workers, REST, Puppeteer or CDP.