One-Line Summary

Big money is flowing to vertical AI while consumer “agent” products go mainstream—raising new security and trust questions as even OpenAI spotlights what orchestrated agents can tackle.

Big Tech

Meta acquires Stilla.ai to expand Business Agent

Meta is buying Stockholm-based Stilla.ai to accelerate development of its Business Agent, which helps businesses complete transactions in WhatsApp, Messenger, and Instagram, according to Axios. Stilla was founded in 2024 and raised $5 million in pre-seed funding; its team and technology are expected to join Meta pending close. ¹

Why it matters: Meta has worked for years to convert chats into purchases, and Axios reports more than 1 million businesses now use its Business Agent to handle customer interactions; the company also plans to expand in Sweden following the deal. For teams exploring chat-to-commerce, this signals more native tooling inside channels customers already use. ¹

Security watch: as agents touch more real data and tools, Meta researchers highlight fresh risks—its “Repeat‑After‑Me” study shows black‑box visual prompt injection can reach attack success rates above 80% on open‑weight models and 47% on commercial frontier VLMs, underscoring the need for strict permissions and isolation. ²

Meta releases Muse personal agent with Secure VM

Wired reports Meta has released Muse, a personal AI agent available via a dedicated iOS/Android app, on the web, and by messaging it directly in WhatsApp. Muse can automate tasks like sending emails and booking travel, and can make purchases using Stripe’s Link; it’s free to try, with paid subscriptions required for heavier automation. ³

Meta is emphasizing privacy and abuse resistance. Wired describes a Secure VM architecture that isolates each user’s agent in a virtual machine and gates actions through a “Sentinel” approval check; a public bug bounty offers up to $300,000, including up to $130,000 for successful prompt‑injection exploits. A “Confidential VM” option with user‑held keys and trusted execution is planned. ³

The New York Times frames Muse as an agent that can send emails and book travel, positioning it within Big Tech’s race to make assistants complete multi‑step tasks end‑to‑end. ⁴

OpenAI claims progress on Navier–Stokes with 10,000 agents

The New York Times reports OpenAI says it has cracked key parts of one of math’s Millennium Problems using thousands of AI agents coordinated on the task; the claim awaits independent verification from the Clay Mathematics Institute. ⁵

BBC adds operational detail: OpenAI describes orchestrating roughly 10,000 bots to reach a result in about 88 hours, exchanging nearly 3 million messages and using about 130 billion output tokens—an effort it estimates would cost around $10 million at list prices. ⁶

BBC also reports controversy: NYU’s Tristan Buckmaster alleges overlapping work and timing concerns; OpenAI denies seeing the researchers’ work, says the proofs differ, and notes it doesn’t intend to claim the Millennium Prize for this result. For non‑research teams, the takeaway is what coordinated agents can do—and how costly such runs can be. ⁶

Industry & Biz

Harvey raises at $15.5B and buys Guardrails AI

Reuters reports legal AI startup Harvey reaches a $15.5 billion valuation in a new funding round, reflecting accelerating enterprise demand for AI that automates legal workflows. ⁷

The Next Web reports Harvey closed a $550 million round at a $15.6 billion valuation and acquired agent‑safety startup Guardrails AI; the round was co‑led by Lightspeed Venture Partners and Diffusion, with Sapphire Ventures and Whale Rock participating. ⁸

TNW adds traction data: annual recurring revenue has surpassed $400 million and the customer base has grown to more than 3,000 organizations—up from 1,300 in March when Harvey was valued at $11 billion after a $200 million raise—with customers including Latham & Watkins and in‑house legal teams at companies such as Microsoft. ⁸

Strategy-wise, TNW reports Harvey is investing in in‑house models—its Tenet is fine‑tuned on Moonshot AI’s Kimi K3—routes tasks to the cheapest capable model, and fine‑tunes on firm documents. Competition is rising from both suppliers (Anthropic plug‑ins, OpenAI partnerships) and vertical peers (Legora). The Guardrails deal aligns with the need to test and govern autonomous legal agents. ⁸

New Tools

Meta Muse: a personal agent you can try

Wired reports Muse is a consumer agent you can message to automate everyday tasks—like sending emails or booking travel—available in a standalone app and in WhatsApp. It’s free to try, with subscriptions needed for heavier automation. ³

Under the hood, Wired describes Secure VM isolation, human‑in‑the‑loop approval prompts, and purchases via Stripe Link’s single‑use cards. A new public bug bounty offers up to $300,000, a signal that security and privacy are first‑order product features. ³

Community Pulse

Hacker News (631↑) — mixed: many distrust Meta’s access to device data, while some early testers praise useful integrations and convenience. ⁹

"How is this expected to work? How is meta going to get access to my data on a phone? Did they build all the integrations needed for every app? This doesn’t look good for privacy, I don’t trust meta at all." — Hacker News ⁹

"I am shocked. After playing with Muse by Meta for an hour in just chat mode, I was impressed enough to: I never use Apple Reminders so I gave Muse read access to Reminders. I asked Muse to schedule repeating tasks that only involve web search e.g., every other morning find sci-fi movies playing in my city, 3pm EST stock market summary, etc. I also permanently set my food preferences and just naming any restaurant in town gets me an accurate list of recommendations with costs and descriptions." — Hacker News ⁹

What This Means for You

Agents are arriving in the apps you already use. With Muse available in WhatsApp and a privacy‑first design (isolation plus approval checks), you can trial a narrow, low‑risk task—like inbox triage or simple travel holds—to see whether the permission prompts and audit trail match your standards. ³

Legal work is becoming a prime AI buyer. Harvey’s reported revenue and customer figures suggest many teams are past pilots and into paid usage, especially for contract review, research summaries, and drafting. If you manage legal workflows, now’s the time to scope a limited RFI with security and data‑handling questions. ⁸

Agent safety isn’t optional. Meta’s own research underscores how prompt‑injection attacks can subvert agents that read from screens or images, so any rollout should include strict app‑by‑app permissions, human‑in‑the‑loop approvals, and red‑team tests before touching sensitive systems. ²

Big achievements in agent orchestration can be expensive. OpenAI’s reported 10,000‑bot run consumed roughly $10 million at list prices—useful perspective when estimating ROI for automation at scale. Start small, automate the repetitive pieces, and reserve heavyweight runs for high‑value cases. ⁶

Action Items

1. Try Meta Muse on a low‑stakes task: Message Muse in WhatsApp or install the app, then assign one narrow job (e.g., draft a travel email) to evaluate approval prompts and results.
2. Run a 30‑minute agent safety check: In Muse, connect only non‑sensitive integrations, attempt a few edge‑case prompts, and confirm you see explicit human‑in‑the‑loop approvals before any action.
3. Request a legal‑AI briefing: Ask your in‑house legal team or primary law firm for a 45‑minute demo of their current AI tools and data‑handling policies; compare against your security checklist.
4. Skim one agent security primer: Read a short guide on agent risks (prompt injection, tool abuse) and update your internal checklist with approval gates and audit‑log requirements.