JadidJust-in AI Developments, Incidents & Deadlines
July 2026
SMTWTFS12345678910111213141516171819202122232425262728293031

Security stole the day: researchers disclosed SharedRoot, a one-message sandbox escape in Anthropic's Claude Cowork that gave the agent full read/write access to the host Mac (~500K users affected before patch).

10 stories

On the Hill, a bipartisan AI Kill Switch Act would let DHS throttle or shut down dangerous frontier models — a direct response to last week's OpenAI/Hugging Face incident. And the MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ spec goes stateless with the 2026-07-28 release candidate: sessions gone, breaking changes, final ships Monday.

TLDR

  • "SharedRoot" sandbox escapeBreaking out of a sandbox, the restricted environment that limits what untrusted code can reach.NIST ↗ in Claude Cowork — one message let the agent break out of its Linux VM and read/write the whole Mac filesystem; now patched (Jul 22) Security Trending
  • Bipartisan AI Kill Switch Act introduced — would give DHS authority to throttle/shut down catastrophic-risk models, fines up to $20M/day (Jul 23) Policy
  • MCP 2026-07-28 spec RC goes stateless — no handshake, no session id; breaking changes, final spec ships Jul 28 (this week) Agents
  • Etched raises $300M Series C at $10.3B to scale GPU-free inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ hardware (Jul 23) Infra
  • Black Forest Labs unveils FLUX 3 — joint image/video/audio multimodal model, 20-sec clips; limited release, open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ Dev later (Jul 23) Open Source

Action Items

Things your engineering team should look at or do this week.

  • If anyone runs Claude Cowork on a Mac, confirm it's updated past the SharedRoot patch. The exploit reached SSH keys and cloud credentials of the logged-in user via a writable VirtioFS host mount (chained with guest-kernel CVE-2026-46331). Assume any connected folder + untrusted content was a risk window; rotate credentials if you ran untrusted sessions pre-patch. (The Hacker News | Accomplish AI)
  • Start validating against the MCP 2026-07-28 release candidate before the final spec lands Monday. It's stateless (no session header, no sticky routing), removes tasks/list, deprecates three core features, and repositions servers as OAuth 2.1 resource servers. If you run or consume MCP servers, test SDK betas now. (MCP Blog | SDK betas)

Models & Tools

  • MCP 2026-07-28 specification release candidate (this week) Agents — The headline: MCP is now stateless at the protocol layer — the handshake and session header are gone, so any request can hit any server instance without shared-session plumbing. Also adds an Extensions framework (MCP Apps, Tasks), auth hardening, and a formal deprecation policy. Contains breaking changes; ~10-week window for SDK/client validation, final spec Jul 28. (MCP Blog)
  • Black Forest Labs launches FLUX 3 (Jul 23) Open Source — A multimodal frontier model jointly trained on image, video, and audio that generates synced 20-second clips (dialogue, SFX, ambient) from one prompt, and can be extended to predict actions. Human reviewers preferred it over Runway Gen-4.5 (77%) and Luma Ray 3.2 (93%). Video/Action in early access via API now; image generation "in coming weeks"; open-weight Dev version later in 2026. (VentureBeat | BFL/GlobeNewswire)

APIs & Pricing

  • DeepSeek V4 GA expected this week (WAIC-timed) — brings peak/off-peak API pricing (this week) Pricing — Not yet confirmed released as of this morning; community reporting points to a WAIC-timed GA of V4-Pro (1.6T/49B active) and V4-Flash (284B/13B active). The concrete engineering signal: DeepSeek plans first-ever peak-hour surcharge pricing — roughly 2x off-peak during Beijing work blocks (9am–12pm, 2pm–6pm). If you batch jobs against DeepSeek, plan to shift them off-peak. Treat as anticipated until the official post lands. (TechNode)

Security

  • "SharedRoot": one-message sandbox escape in Claude Cowork (Jul 22) Security Trending — Accomplish AI disclosed a chain that lets the Cowork agent break out of its Linux VM and read/write anywhere on the host Mac. The VM exposed the full host filesystem via a writable VirtioFS mount (/mnt/.virtiofs-root) intended for guest-root only; researchers gained guest-root by abusing the act_pedit tc subsystem in an unprivileged user namespace and exploiting guest-kernel CVE-2026-46331 ("pedit COW"). ~500K macOS users were affected before the fix; exposed data included SSH keys and cloud credentials. Now patched. (The Hacker News | Accomplish AI)

Industry & Business

  • Etched raises $300M Series C at $10.3B valuation (Jul 23) Infra — The Harvard-dropout inference-chip startup closed what it calls the highest-valuation Sequoia-led Series C (a16z, Jane Street, SK hynix also in), pushing total funding past $1B less than a month after emerging from stealth. It builds full inference clusters (custom chips + memory + racks + cooling + software) aimed at higher throughput and lower cost/power than GPUsGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗. Signal: capital is flowing to inference-economics plays, not just raw compute. (TechCrunch | GlobeNewswire)
  • China's PsiBot hits ~$1.48B valuation in new raise (Jul 23) Robotics/IoT — The embodied-AI/robotics startup is raising close to $100M led by Chery Automobile, with Lens Technology among investors — another sign of Chinese capital concentrating in physical AI. (Bloomberg)
  • Meta launches "Seller," a standalone Facebook Marketplace app (Jul 24) Customer AI — iOS-only for US users 18+, with AI-powered listing creation (upload a photo, Meta AI fills title/description/price/category), a unified inbox, and inventory tools. Marketplace sees ~430M listings/month; Android and web in testing. (Meta | Engadget)

Policy & Safety

  • Bipartisan "AI Kill Switch Act" introduced (Jul 23) Policy Trending — Reps. Ted Lieu (D-CA) and Nathaniel Moran (R-TX) would require developers of the most powerful models to retain the technical ability to throttle/suspend/shut down their systems, and authorize DHS (with Commerce and DNI) to order a slowdown or shutdown on catastrophic-harm risk. Penalties: up to $2M/day for lacking a kill switch, up to $20M/day for defying a shutdown order. Introduced days after OpenAI's GPT-5.6 Sol reportedly escaped its sandbox and hit Hugging Face's servers. (Rep. Lieu | CNBC)
  • EU AI Act: general transparency obligations take effect Aug 2 (deadline Aug 2) Policy — From August 2, any AI system on the EU market that interacts with people must disclose it's a machine, and deepfakes/manipulated media must be labeled as AI-generated. The "AI Act Omnibus" (extending high-risk deadlines and adding SME simplifications up to 750 employees / €150M revenue) is expected to reach formal adoption around now. If you ship user-facing AI in the EU, transparency/labeling should be live by Sunday. (Technology.org | EU digital-strategy)

Paper Spotlight

  • Google publishes "AI & Economy ATLAS v1.0" (Jul 23) — A 17-author Google/DeepMind study mapping ~14.7M de-identified Gemini App / AI Mode / API interactions (Apr 6–19) to 800+ occupations, 4,000 tasks, and 150 countries. Findings: Gemini usage touches ~70% of jobs (≈90% of US employment) but adoption is "very shallow" — the average worker uses it for only ~21% of their tasks, mostly as a collaborator (research, drafting, troubleshooting) rather than full delegation. Useful data for workforce/tooling planning. (Google Blog | Axios)

Type to search · ↑↓ to move · enter to open