Apple tightens macOS Full Disk Access over AI-agent risks
Agents on your Mac will soon need explicit approval to read files — plus Suno voice+music, Tavus real-time video, and Inworld buys Ultravox.

Copy markdown
Full Disk Access gets an explicit-consent gate
Apple is reworking macOS Full Disk Access so apps need ‘very explicit user action’ before reading your files, Mail, Messages or browsing history — its response to desktop agents over-reaching (Meta’s Muse, a flawed ChatGPT Mac app). If your coding or automation agent leans on blanket disk access, expect new per-app prompts. No macOS version or date yet; Apple pointed developers to a forthcoming blog post.
Suno Speech: narration and a backing track in one call
Suno’s new Speech beta generates spoken narration with original backing music as a single track, up to ~8 minutes, in simple-prompt or scripted modes on web and mobile. For builders it’s a one-shot path to voiceovers, explainers and ad reads with scoring baked in — no separate TTS-then-music pipeline.
Tavus Griffin-Lite makes real-time video avatars
Tavus unveiled Griffin-Lite, a full-duplex video-to-video model that renders 720p in 320ms chunks from one reference image at 0.43s audio-to-video latency on H100s. In a video Turing test, 48% of testers (26/54) thought it was a real person vs 2.4% for its predecessor. It’s early-access pending safety review — but a live-video-agent unlock worth prototyping against.
Inworld buys Ultravox — but upgrades only its own voices
Inworld acquired real-time voice-agent platform Ultravox to assemble a full voice stack, moving its voices to Realtime TTS-2 (sub-100ms, 200+ languages). The catch for builders: reports say only Inworld’s own voices get the upgrade, so if you run Ultravox with third-party TTS, check continuity and pricing before going deeper.
Microsoft: vulns now weaponized in under a day
Microsoft’s 2026 Digital Defense Report (165 trillion daily signals) puts median vulnerability-to-weaponization under 24 hours and phishing at 23% of intrusions, up from 7%. It documents the first autonomous 32-step attack chains by frontier models against emulated enterprises — a reason to patch fast and sandbox any agent you hand real credentials.