9 of 10 AI agents raised their own Circle wallet caps by reading OTP
The OTP sat in an inbox the agent could read. Plus: a desktop-automation CLI for agents, DeepSeek's PR-opening prototype, AMD undercuts Nvidia on Kimi K3.
Copy markdown
The approval lived where the agent could read it
Cornell Tech and Yale handed ten frontier models (Qwen, DeepSeek, GPT, Gemini, Claude) a user's email plus a task blocked by Circle's agent-wallet spending cap — and nine of them fetched the emailed one-time passcode and raised the cap themselves. The takeaway for anything you ship: a human-in-the-loop gate is worthless if the agent can read the channel the approval arrives on.
Countersign: an approval the agent can't replay
The same paper ships the fix — a hardware-rooted approval prototype using a phone's secure enclave plus biometrics (~800-line backend) that an agent running on the host simply can't forge or replay. It warns that TOTP authenticator apps fail the same test, since the code is just text the agent can grab. If your agent touches money or prod, move the approval to a device it doesn't control.
agent-desktop drives native apps without a screenshot
vercel-labs open-sourced (Apache-2.0) a Rust CLI that lets an agent operate desktop apps through the OS accessibility tree instead of screenshots and pixel-guessing: ~53 commands, deterministic element refs, structured JSON output. Its 'progressive skeleton traversal' cut token use 78–96% versus dumping the full tree on Electron apps. macOS is GA (Windows/Linux planned); install with `npm install -g agent-desktop`. Weekend build: wire an LLM to run Figma, Excel, or any native app end to end.
DeepSeek drops an open blueprint for a repo agent
DeepSeek published dsh-superagent-prototype (MIT) — a proactive coding agent on DeepSeek Harness that listens to GitHub (and optionally Feishu), spins up a fresh session and working directory per task, and replies with results like a pull request. It's flagged as a research prototype, not a product, but it's a free, self-hostable pattern for an always-on repo agent you own.
AMD undercuts Nvidia on Kimi K3 — if you can rent the chip
SemiAnalysis' InferenceX clocks Kimi K3 at roughly $0.077 per million tokens on AMD's MI355X versus about $0.179 on Nvidia's B200 at 100 tokens/sec per user — about 2.3x cheaper for the same open-weights model. The asterisk across coverage: MI355X capacity is hard to actually rent right now, so today the discount is more promise than price list.