Least-privilege permissions for AI agents

Jev 1.14, the System One decision model, is live under the jev-preview alias with three experimental 'danger zone' features: images, a box type and longer context. Context is 64k tokens per request by default, or 262k with an opt-in header; TypeSafe says none of the three features are production-ready.

AI Search queries agent traces in natural language ("did the user get frustrated?", "did PII leak?") instead of filters or pre-built evaluators. Engine v2 adds red teaming and auto-validates proposed fixes; Managed Deep Agents v0.9 adds user-level memory, Slack, scheduling and free web search.

Krea 3 combines separate references for characters, outfits, and locations into a single scene while keeping faces and color grading consistent. It is benchmarked on CPI-Bench.

GPT-6.1 Sol reviewed nearly 800 mathematics papers, flagging major errors in 6% and minor mistakes in another 30%; researchers estimate 90% of its major findings are correct. The major-error rate in newer papers fell to 1.5% as AI adoption surged.

Since its release last month, Meta's Muse agent has been downloaded millions of times and hit No. 1 on Apple's App Store while requesting access to bank accounts, passwords, contacts, and chat logs. Signal Foundation president Meredith Whittaker told The Verge the agent acts on everyone a user communicates with, not just the installer.

Two Reddit users say OpenAI deactivated their paid accounts citing "election interference" and "Recidivism" respectively; both deny the violations and have appealed. One reports losing half a month of a paid subscription.