Welcome to the daily AI news brief for vibe coders. Today is Thursday, June 25, 2026. The theme is agents moving into shared workspaces and tightening their guardrails: Claude now sits inside Slack channels, Claude Code locks down secrets and hung MCP calls, and Google’s Imagen image models are gone. If you build on Shopify, Claude, GPT, Gemini, or Firebase, here is what moved in the last 24 hours and what to do about it.
TL;DR
- Claude Tag puts a shared, async Claude agent inside Slack channels for Team and Enterprise.
- Claude Code adds sandbox credential blocking, org model restrictions, and a 5-minute MCP hang abort.
- Google Imagen image models retired yesterday; confirm your Gemini “Nano Banana” migration.
- ChatGPT now converts pastes over 10,000 characters into attachments for Free and Go users.
- Anthropic publicly accused Alibaba of illicitly extracting its model capabilities.
- Build of the day: harden one Claude Code project against secret leaks and hung MCP calls.
Claude Tag puts a shared agent in Slack [STACK]
What shipped. Anthropic launched Claude Tag, a Slack-resident async agent, as a research preview for Claude Team and Enterprise customers. You @Claude a request in a channel, the agent breaks it into milestones, runs them with its connected tools, and posts results back in-thread. The differentiator versus older Slack bots is a single shared Claude per channel with shared memory, plus an ambient mode that can post unprompted. It runs on Claude Opus 4.8.
Why it matters for vibe coders. A shared agent identity lets teammates hand off half-finished tasks to the same Claude, which is closer to a multi-agent workflow than a personal chatbot.
What to do today. If you are on Team or Enterprise, scope Claude Tag to one low-risk channel and watch how it sequences a multi-step task before granting broader tool access. The New Stack
Claude Code hardens secrets and MCP [STACK]
What shipped. The latest Claude Code update added a sandbox.credentials setting that blocks sandboxed commands from reading credential files and secret environment variables, plus org-configured model restrictions across the model picker, --model, /model, and ANTHROPIC_MODEL. Remote MCP tool calls that hang with no response now abort with an error after five minutes instead of blocking indefinitely, and fullscreen mode gained mouse-click support for select menus.
Why it matters for vibe coders. If you run agents against your own machine, a sandbox that cannot read secrets and an MCP loop that fails fast remove two of the quietest ways an agentic build can stall or leak.
What to do today. Set sandbox.credentials in your highest-risk project and re-run an agent task that touches MCP to confirm it errors cleanly instead of hanging. Claude Code Docs
Google Imagen is retired — verify your migration [STACK]
What shipped. Google shut down all Imagen image models on the Gemini Developer API and Vertex AI on June 24, 2026. Calls to retired imagen-* endpoints now return errors. The named replacement is the Gemini Image family, code-named “Nano Banana”; the primary change is swapping the model string, after which request and response formats are largely compatible.
Why it matters for vibe coders. If your storefront, product mockups, or design tooling generate images through Firebase or the Gemini API, a hard-coded Imagen model is broken as of yesterday, and a blank product image is easy to ship unnoticed.
What to do today. Grep your code and Firebase configs for “imagen,” swap to a supported Gemini image model, and regenerate one asset end to end to confirm the path works. Firebase
ChatGPT turns big pastes into attachments [ECOSYSTEM]
What shipped. ChatGPT now handles large pastes as attachments for Free and Go users, automatically converting any paste over 10,000 characters into an attachment to keep the composer clean and protect context. Users can move the content back into the message at any time with “Show in text field.”
Why it matters for vibe coders. If you paste long logs or files to draft prompts, the attachment behavior changes how much text lands directly in context, which affects how the model reads your input.
What to do today. Paste a long error log into ChatGPT and decide, per task, whether to keep it as an attachment or expand it back into the message before you rely on the output. OpenAI
Also worth noting
- Anthropic publicly accused Alibaba of a campaign to “brazenly” and illicitly extract its AI model capabilities, a reminder that model-distillation disputes are now front-page business news. CNBC
- Documents show Meta plans to launch an AI-powered prediction market app to compete with Kalshi and Polymarket, signaling more consumer AI products colliding with regulated markets. KCBX/NPR
Build of the day
Spend two hours hardening one Claude Code project against the two failure modes in today’s news. First, add a sandbox.credentials setting so sandboxed commands cannot read your credential files or secret environment variables, then run an agent task and confirm it cannot print a secret. Second, point the agent at a remote MCP server and trigger a long-running tool call to confirm it now aborts cleanly after five minutes rather than hanging. Log which guardrail fired. You finish with a project that fails fast and keeps secrets out of agent output.
FAQ
What is Claude Tag and who can use it?
Claude Tag is a Slack-resident async agent from Anthropic, launched as a research preview for Claude Team and Enterprise customers. You @Claude a request in a channel; the agent breaks it into milestones, runs connected tools, and posts results in-thread. It uses a single shared Claude per channel with shared memory rather than a per-user bot. The New Stack
Does Claude Tag run on Claude Opus 4.8?
Yes. Claude Tag runs on Claude Opus 4.8, Anthropic’s most capable publicly available model. The shared-agent design means everyone in a channel collaborates with the same Claude identity and can hand off in-progress tasks, which is closer to a team workflow than a personal assistant. Scope it to one channel before granting broad tool access. The New Stack
What did Claude Code change for MCP and secrets?
The latest Claude Code update added a sandbox.credentials setting that blocks sandboxed commands from reading credential files and secret environment variables, plus org-configured model restrictions. Remote MCP tool calls that hang now abort with an error after five minutes instead of blocking indefinitely. Update Claude Code and re-test an MCP task to confirm the new behavior. Claude Code Docs
Do I still need to migrate off Google Imagen?
Yes, and the deadline has passed. Google shut down all Imagen image models on the Gemini Developer API and Vertex AI on June 24, 2026, so imagen-* calls now return errors. Migrate to a supported Gemini Image (“Nano Banana”) model by swapping the model string, then regenerate one asset end to end to verify output. Firebase
How does ChatGPT handle large pastes now?
ChatGPT now converts any paste over 10,000 characters into an attachment for Free and Go users, keeping the composer clean and protecting context. The content is not lost; you can click “Show in text field” to expand it back into the message. Decide per task whether long logs belong in the message body or as an attachment before relying on the answer. OpenAI
Why does the Anthropic and Alibaba dispute matter to builders?
Anthropic publicly accused Alibaba of a campaign to illicitly extract its AI model capabilities, the kind of distillation dispute that can shape model access terms and licensing. For builders, it is a reminder to read the usage terms of any model you depend on, since provider policies and availability can shift when these disputes escalate. CNBC
Sources
- The New Stack — Anthropic gives @Claude a seat in Slack
- Claude Code Docs — What’s new
- Firebase — Migrate from Imagen to a Gemini Image model
- OpenAI — ChatGPT release notes
- CNBC — Anthropic accuses Alibaba of distillation campaign
- KCBX/NPR — Meta plans an AI-powered prediction market app
About the author
Robert McCullock builds AI-assisted commerce on Shopify, Claude, and Firebase at Design Delight Studio. See his background and project work in his professional portfolio.