The Product Desk

Trail of Bits gave GPT 5.6-Cyber one objective, escape a QEMU/KVM virtual machine on Debian 12, and it broke out three times across roughly 12 unsupervised hours, including once against a fully patched host. Only AWS Firecracker, a stripped-down microVM, held. Checkmarx separately demonstrated that human-in-the-loop approval prompts can be forged by the agent they gate. Both controls are load-bearing in most agent specs and trust pages written in the last 18 months, which makes those documents inaccurate today.

Duckbill Group cut median merge time from 26 hours to 1 hour and merged 94% more pull requests after dropping human review on non-critical paths, per Gergely Orosz's reporting. The price of admission was an 85% unit-test coverage floor plus near-total lint rule enforcement, so review labor converted into token spend. Their human-review-required list is four-fifths product surface: public API/MCP, auth, design system, non-additive schema changes, agent skills. Pricing, billing, entitlements and the onboarding funnel are missing from it.

OpenDesign turns Harness on by default for every user tomorrow, September 9, claiming state-of-the-art design quality at roughly one-twentieth its prior cost. The output is not a first-screen mockup: complete clickable apps with real navigation, multi-step flows and state that persists across pages, built from a brand's own brief and design system. The only quality evidence is one survey with more than 88% positive responses and no disclosed methodology. For you, the artifact a usability test needs becomes generated rather than built, and the discovery constraint moves from design bandwidth to participant recruiting.

Claude Fable 5.1 is now available through Amazon Bedrock with prompts and outputs subject to retention for up to 30 days and to possible safety review by Amazon personnel. The remedy, Anthropic's Enterprise Frontier Safeguards, is unreleased and limited to eligible customers. Cursor shipped the complementary pattern the same day: customers run agent workers on their own machines across AWS, Vercel, Modal, Cloudflare or Macs while Cursor keeps the agent loop and inference. Any feature you ship on Bedrock inherits that retention window in your next security questionnaire.

Hugging Face co-founder Thomas Wolf says Microduck has sold 15,000+ units and over $6 million in about a month, with advertising confined to X and LinkedIn. Reachy Mini sold roughly 10,000 units over its entire lifetime at the same $400 price, same brand, same channel, same audience. The isolated variable is capability: Microduck walks and acquires skills, while its predecessor saw, listened and spoke. Read as pricing rather than robotics, a deliberately cheap experimentation tier priced near 1% of a $40,000 prototype is what sold.