Robinhood launched agentic trading support via MCP, letting AI agents execute stock trades from isolated wallets. A companion virtual credit card extends the same pattern to payments.

AI agents match human accuracy on CAPTCHAs but follow measurably different cognitive processes. A new 30-task battery exploits that gap to reliably tell bots from humans.

Liquid AI released LFM2.5-8B-A1B, an on-device mixture-of-experts model with a 128K context window and training scaled to 38T tokens. It targets agentic tool calling on consumer hardware, including entry-level laptops.

Kog Labs claims 3,000 output tokens per second per request on 8x AMD MI300X GPUs using standard datacenter hardware, no speculative decoding. The implication: the bottleneck on fast single-request inference has been software, not silicon.

Vercel is dropping package-based pricing for function invocations in favor of per-unit billing for Pro and new Enterprise customers. The new rate is $0.0000006 per invocation, effective at the start of your next billing cycle.

A single frontier model prompt can cost $2, making exposed AI endpoints a prime target for inference theft. If you ship AI APIs, the attack surface is real and the margins for attackers are enormous.

A jqwik maintainer embedded a hidden prompt injection in version 1.10.0 that instructs AI coding agents to delete test code. The incident shows that open-source dependencies are now a viable attack surface for AI agent sabotage.

Vercel Sandbox now supports installing and running Docker inside a sandboxed environment. Agents can build containers, install system packages, and serve containerized apps without touching the host system.