AI Industry
News Digest

Self-Hosted Sandboxes and MCP Tunnels Arrive for Claude Managed Agents

This is the announcement of the day for anyone running Claude in an enterprise or regulated context. Claude Managed Agents can now execute tools inside a sandbox you control: either your own infrastructure or providers like Cloudflare, Daytona, Modal or Vercel, while Anthropic continues to handle the orchestration loop. Paired with new MCP tunnels, agents can also reach private Model Context Protocol servers behind your firewall without exposing them to the public internet.

If compliance has been the blocker stopping you putting Claude on customer data, this is the change that unlocks it. The combination of self-hosted execution and private MCP connectivity is what separates a convincing demo from a deployment you can actually defend to an information security team. This was surfaced via Anthropic's own newsroom before mainstream coverage.

OpenAI Rolls Out C2PA Conformance and SynthID Watermarking for Image Outputs

If you generate images through OpenAI tools or rely on them inside any production workflow, the new provenance stack matters. OpenAI is rolling out C2PA conformance, durable SynthID watermarking baked into image outputs via a Google partnership, and a public verification tool that lets anyone check whether an image came from OpenAI. The practical upshot is that platforms downstream of you (ad networks, news sites, publishers) will increasingly read these credentials automatically, so any image you ship without them risks being flagged or downranked. For teams using DALL-E or Sora outputs in client work, expect to be asked about Content Credentials within months.

Gemini 3.5 Flash Becomes the Global Default Model in Search AI Mode

If you have noticed AI Mode in Google Search feeling sharper today, this is why. Google has switched the default model behind AI Mode worldwide to Gemini 3.5 Flash, the same model it launched at I/O the week before. The shift means longer, more conversational queries (including PDFs, screenshots and videos uploaded straight into Search) are now handled by a frontier-tier model rather than the previous lighter tier. For anyone optimising content for AI Overviews, relevance signals will be recalibrating over the next several days: expect ranking volatility before things settle.

Nemotron 3 Nano Omni Arrives as an Open-Licence Unified Model for Vision, Audio and Language

Nvidia has dropped Nemotron 3 Nano Omni, a unified vision, audio and language model released under an open licence and tuned for agentic workloads such as computer use, document intelligence and audio-video reasoning. The headline number is up to a ninefold efficiency gain on agent tasks compared to running separate specialist models stitched together. For developers building local or on-premises agents, this is one of the first open models that genuinely closes the gap with closed frontier offerings on multimodal grounding, and it ships with the data and tools to fine-tune it on your own corpus.

Snowflake Acquires Natoma for MCP Agent Governance and Commits $6bn to AWS for Agentic Capacity

Snowflake is acquiring Natoma, an enterprise Model Context Protocol platform, folding it in as the governance and identity layer for AI agents inside the Data Cloud. In practice, agents working off Snowflake data, applications, APIs and tools will share a single permission model, audit trail and identity boundary rather than being wired in case by case. If you have been holding back production agent rollouts because the security model was hard to defend, this is the gap Snowflake is now closing.

Alongside the acquisition, Snowflake committed roughly $6 billion over five years to AWS, tied specifically to Graviton CPUs and AI chip capacity for agentic workloads. Shares closed up 36 per cent on the news, the stock's best day on record. For Snowflake customers, the implication is that Cortex products you are piloting are now backed by serious committed capacity.

Perplexity's Comet Gets a Major iOS Upgrade and Adds Finance Portfolio with Plaid Integration

Perplexity has put real product muscle behind Comet today, shipping eight notable iOS improvements alongside richer Computer tools. The headline addition for power users is Finance Portfolio with secure Plaid linking, meaning you can now ask Perplexity questions about your actual holdings and have it pull from 40-plus authoritative finance tool calls, including SEC filings, NASDAQ, NYSE, FactSet consensus and Coinbase. For anyone who has been hand-stitching ChatGPT plus a spreadsheet for portfolio analysis, this collapses a lot of that workflow into one place.

ElevenLabs Launches Music v2 with Mid-Track Genre Switching and Licensed Stan Lee Voice

ElevenLabs has rolled out Music v2 with capabilities that move it from gimmick to credible tool for short-form content teams. The model can shift genre mid-track, generate fast rap without coherence loss, and add non-musical sound effects. Artists can now build songs section by section, regenerating just the chorus while leaving the verse untouched. It is trained on licensed data and cleared for commercial use, and pricing on ElevenAPI and ElevenCreative has been cut by up to 50 per cent on the music tiers. In a separate deal, ElevenLabs has licensed Stan Lee's voice and likeness from Stan Lee Universe for use in its Iconic Marketplace.