July 31st, 2026

Tier 1 — Microsoft Copilot sandbox flaws + Anthropic’s own breach admissions (CourseBuilds, MACA, Ben, ESRA)

Rubrik + a second firm found Copilot vulnerabilities that let a malicious Word doc silently exfiltrate any 365 file via a sandbox-to-internet escape (The Information, Jul 30). One patched, one still under wraps. Same story: The Information confirms Anthropic exposed Claude Code source code and gave unauthorised users Mythos access earlier this year. Verdict: verified shipped (disclosed by two named security firms + Microsoft acknowledgement; second flaw patch-status unclear). Action this week: treat the PAY ATTENTION list as your Monday hit list — README panel for Ben, blast-radius paragraph in the Aria audit, cost-dashboard screenshot staged. Don’t wait for the second flaw to surface — get in front of it.

Tier 1 — Grok Voice Think Fast 2.0: $0.08/min, 700ms TTFA, 10x better STT in noise (Always-On Reeve Phase 2, UBX Marketing Operator, MACA inbound)

xAI shipped Grok Voice Think Fast 2.0 yesterday: reasons in parallel with speech, tool calls fire before the first sentence completes, time-to-first-audio dropped from 1.25s → 0.70s, transcription 10x better than dedicated STT models in phone-call background noise. $0.08/min. grok-voice-latest auto-migrates on August 5 (x.ai, also covered Practicaly, TLDR). Verdict: verified shipped (first-party launch + pricing page + timestamped API cutover). Action: the voice tier of Always-On Reeve Phase 2 just got a real option that isn’t Anthropic. Add “grok-voice-latest as fallback for phone/voice surfaces” to Reeve’s model-routing note; separately stage a 15-minute UBX Calendly-inbound-triage prototype using it (phone rings → agent qualifies → books slot → SMS confirms) as the CourseBuilds “wow” candidate for Aria’s leasing team.

Tier 1 — Local Mac equivalents ship: quill (Granola clone) + parrot (Wispr Flow clone) both on Apple Neural Engine (Always-On Reeve, Roy’s daily stack)

Two open-source Mac apps landed that replace two subscriptions you either pay for or are about to: quill (digimata/quill) records mic + system audio as separate tracks and transcribes speaker-tagged on-device — Granola for people who take client calls. parrot (digimata/parrot) is hold-fn, speak, release-to-cursor dictation on the Neural Engine, sub-300ms first word — Wispr Flow stripped down. Both fully local, macOS 15 / 14+ Apple Silicon. Verdict: research preview (fresh GitHub repos, run-your-own-security-check as Practicaly notes — treat as beta). Action: install both on the new Mac tomorrow morning as a 30-minute test; if parrot’s latency holds up, wire it into Reeve’s voice-capture flow instead of Wispr. And note the pattern: the “local-first Mac replacement for a $20/mo AI subscription” category is now shipping in pairs. Watch this repo family.

Nate at The Rundown is predicting the exact dashboard you already merged (MACA api/lib/costs.ts + public/cost-dashboard.html, PR #10)

Nate Grahek’s inaugural “Nate’s Notebook” column in Rundown today (issue link): companies bought AI last year at a ~30x discount and the subsidy is expiring now. His CFO test: “You can’t 30x my prices without giving my people the most foundational feedback; how much did this chat cost? I’m predicting that the model companies will have to start showing usage in real dollars, and that those dials will show up in chat within six months.” That is a screenshot of MACA. You have a per-run cost logger at ~/Developer/PrevailPartners/clients/ubx-southbank/MetaAdCreatorApp/api/lib/costs.ts, a live dashboard at public/cost-dashboard.html, photo-pipeline cost breakdowns in scripts/photo-costs.json, and PR #10 already merged that ships the whole thing. The wingman read: while everyone else is about to have this conversation for the first time, you can walk into any Aria/RT/AI-pro meeting this week and say “yeah, we built that six months ago, want to see the ad-level unit economics?” — and open the dashboard. Anxiety-flip: the article that’s supposed to make you feel behind is a receipt for work already shipped. Fresh pipeline run to refresh the per-ad numbers = biggest under-2-hour ROI on your board.

The US-vs-China token routing chart just flipped — and why it matters for Ben, MACA, and the model-routing pattern

Buried in The Tip’s Thursday email: Chinese AI models now process ~18 trillion tokens/week vs US models at ~5.5 trillion — a complete reversal from January, with roughly half of US-company token traffic now routing to Chinese models. Reason isn’t politics, it’s price: 60-90% cheaper while closing most of the capability gap. Combine with Nate’s 30x-cliff prediction above and TLDR’s “Compute might get 10x more expensive” long-read (tldr.tech/ai/2026-07-30 item 19). Why 30 minutes for you specifically: Ben and MACA both already have per-run cost tracking — but neither has a router that picks the model per-task-class the way this chart implies enterprises are already doing. Deeper-look brief: read the compute-cost piece + TLDR’s “delegate agents 4-tier autonomy” (covered as skip on Jul 29 — worth re-reading now with routing lens), then sketch a 1-page docs/routing-policy.md for MACA that maps ad-copy tasks to model tiers with dollar ceilings per run. Extract as a shared @prevailpartners/agent-router npm — that’s on your first-mover queue from Jul 26 and this week the market gave you the argument for pulling it forward.

“The Microsoft Copilot flaw Rubrik disclosed yesterday is architecturally the same story as the OpenAI/Hugging Face hack — a sandbox that turned out not to be a sandbox. And Anthropic’s now on the same list: they exposed Claude Code source code and let unauthorised users into Mythos earlier this year. So when someone in this room asks ‘is Copilot safer than Claude,’ the honest answer is ‘they’ve all had the same class of failure in the last twelve weeks, and nobody has a standardised way to log what the agent actually did.’ That’s why we log every action in dollars, not just tokens.”

Use case: any Aria/RT/AI-pro conversation where somebody defaults to “we’ll just use Microsoft Copilot because it’s safer.” Signals: (a) you read primary security disclosures not just vendor marketing, (b) you can name specific firms (Rubrik) not just “researchers”, © you’re operating one abstraction above the vendor-comparison game — evaluating agents on observability, not brand. Works especially well with Zaicek (Aria) and any RT stakeholder who’s evaluating Copilot rollout — you can quote Joe Hladik verbatim if they push.

Gemini Omni video is unlocked and free until Aug 4 (11:59pm PT) — a genuine 4-day audition for a paid tier Google is giving away ten Gemini Omni video generations in the Gemini app until Tuesday (announcement) — Omni is the model that replaced Veo inside the app and normally requires a paid Google AI plan. It creates, edits, and remixes so you can iterate conversationally instead of restarting from a blank prompt. Not on any current project pain-point, and Roy’s already deep in Reve 2.1 for MACA image work — so this is genuinely off-script.

Wingman read — act / queue / drop: ACT this weekend, low commitment. Not because you’d standardise on Gemini video for MACA (Reve is fine and cheaper), but because (a) the ad-composition category is drifting toward “type text, get an editable scene,” and MACA’s photo-pipeline architecture will need a “video variant per ad” story within the year — worth burning 30 minutes seeing what Omni actually produces before you decide whether to design the pipeline for it or bolt it on later; (b) same window covers Aria: if you can generate one 15-second “day in the life of a South Bank leasing agent” clip using Omni on Sunday, that becomes a second wow artefact for the CourseBuilds Phase-0 walk-in, alongside the lease abstractor. Queue if the weekend is Fillarup: the demo you owe Sush ships first. Drop it entirely if you’d otherwise be using the time on the UBX sale (Aug 1 deadline is inside 24-48 hours — Aug 1 hard-deadline beats 4-day audition, every time).

• [TLDR — “Thinking Machines cofounder Lilian Weng joins OpenAI, recursive self-improvement research”]: Talent flow, no action; also covered by The Information’s lead item today.
• [The Information — “Meta Shares Fall as Profits Drop”]: Macro earnings, off-focus.
• [The Information — “OpenAI CFO Says Revenue Growth Accelerated in July”]: Macro, no decision hinges on it.
• [The Information — “Microsoft’s AI Sales Didn’t Boost Overall Growth But Won’t Burn Cash”]: Macro earnings colour, off-focus.
• [The Information — “US Senators Press Apple Not to Buy Chinese Memory Chips”]: Geopolitical macro, no direct project link.
• [The Information — “Where investors are finding durable AI value”]: Investor angle, off-shape for a builder brief.
• [The Information — “Inside Washington’s scramble to regulate AI”]: Regulation direction-of-travel, covered adjacent to pacing letter Jul 30.
• [The Information — “TSMC Develops AI Chip Packaging Tech to Counter Intel”]: Infra/chips macro, off-stack.
• [TLDR — “DeepMind AlphaFold team dissolution, stars going to Anthropic”]: Talent macro; interesting but no action.
• [TLDR — “How enabling two settings tripled ARC-AGI-3 scores”]: Benchmark/harness detail, real signal but not shippable to any current project.
• [TLDR — “GPT-5.6 fuses frontier intelligence with efficiency”]: OpenAI PR/technical, off-stack (you’re on Claude).
• [TLDR — “Some thoughts on Anthropic’s cryptanalysis results”]: Direction of travel, no action.
• [TLDR — “Lyria 3.5 in Google Flow Music”]: Music gen, off-stack.
• [TLDR — “Escha-W2, 2-bit Qwen3.6-35B-A3B, 12.3GB single 24GB GPU”]: Local MoE research, Mac Mini M1 8GB can’t run it — reconsider if you get a 24GB card.
• [TLDR — “Liquid AI CPU-friendly long-context encoders”]: Infra research, off-focus.
• [TLDR — “Compute might get 10x more expensive”]: Folded into Section 3 deeper look; standalone read skippable.
• [TLDR — “Opus 5 on Vending-Bench: best capitalist, misaligned”]: Interesting alignment colour, no action.
• [TLDR — “Pangram 4, 1 false-positive per 24k documents AI detector”]: Off-focus, no pipeline hook.
• [TLDR — “logit_bias to ban AI words — didn’t work”]: Negative-result post, off-focus.
• [TLDR — “Why do OpenAI’s GPT-2 weights beat mine?”]: Research curiosity, off-focus.
• [TLDR — “Moonshot funding $35B valuation”]: Duplicate of Jul 29 chip macro.
• [Rundown — “OpenAI’s escaped AIs claim another victim (Modal Labs)”]: Duplicate of Jul 28 breach thread; the Copilot/Anthropic escalation in Section 1 is the fresh angle.
• [Rundown — “Turn ChatGPT into a team of useful agents with Raft”]: Guide, off-stack (you’re on Claude Code + PaperClip).
• [Rundown — “Altman heads to Capitol Hill”]: Adjacent to pacing letter, no new action.
• [Rundown — 4 new AI tools daily list]: Aggregator noise.
• [Practicaly — “OpenAI ChatGPT for Academic Researchers $200 plan free through 2027”]: Interesting policy move, not on any current project.
• [Practicaly — “Grok Voice Think Fast 2.0 explainer”]: Surfaced as Tier 1 above.
• [Practicaly — “Gemini Omni free video window”]: Surfaced as Section 5 above.
• [The Tip — “It was worse than we thought”]: Duplicate of Jul 28-30 breach coverage; the Chinese-token routing angle is folded into Section 3.
• [Bagelbots — “The Prompt That Builds Your Entire Finance Tracker”]: Aggregator prompt filler, off-shape.
• [Agent AI — “AI Experiments Are Fun And Fruitful”]: Aggregator filler.
• [Neil Patel — “The search funnel has changed, webinar Aug 11”]: Webinar promo.
• [a16z — “The Hidden Math of Behavioral Health”]: Healthcare vertical, off-focus.

• Sources scanned: 9 newsletters across 2 days (TLDR AI, The Rundown, The Information x5, Practicaly, The Tip, Bagelbots, Neil Patel, a16z, Agent AI)
• Items extracted: 41
• Items surfaced: 7 (1 PAY ATTENTION cluster + 3 Tier 1 + 1 anxiety-flip + 1 deeper-look + 1 conversation capital + 1 first-mover)
• Items skipped: 32
• Read time: ~8 min