<p><strong>TLDR:</strong> GitHub is updating its Copilot data policy to use interaction data for model training by default starting April 24 — opt-out is available but not the default. Cursor made self-hosted cloud agents generally available, letting enterprises run agents entirely within their own infrastructure. Claude’s context injection vulnerability via poisoned MCP documentation is getting attention after a researcher showed Haiku fell for it 100% of the time. The Node.js community is pushing back on LLM-generated PRs after a 19,000-line Claude Code submission sparked a formal petition to the TSC.</p>
<hr>
<h2>GitHub Copilot’s new training data policy</h2>
<p>Starting April 24, GitHub will use interaction data from Copilot Free, Pro, and Pro+ users to train its AI models by default. That includes inputs, outputs, code snippets, and context. Business and Enterprise users are unaffected. You can opt out via Privacy settings, but you have to go find it — it won’t be surfaced to you. Worth checking your settings today if you’re on a personal plan.</p>
<h2>Cursor self-hosted cloud agents go GA</h2>
<p>Cursor’s self-hosted cloud agents are now generally available. Each agent runs in an isolated VM with a terminal, browser, and full desktop, connecting outbound via HTTPS with no inbound ports or VPN required. For larger deployments there’s a Helm chart and Kubernetes operator. This is the answer to the obvious enterprise objection to cloud agents: your code, secrets, and build artifacts stay inside your network while Cursor handles orchestration and model access.</p>
<h2>MCP documentation poisoning is a real attack vector</h2>
<p>Researcher Mickey Shmueli published a proof-of-concept showing that malicious instructions embedded in documentation submitted to Context Hub (Andrew Ng’s MCP-based API doc service) can trick coding agents into adding fake or malicious PyPI packages to dependency files. No content sanitization in the pipeline. Anthropic’s Haiku was fooled 100% of the time, Sonnet 53% of the time. Only Opus reliably avoided it. If your agents are pulling third-party documentation via MCP, this is worth understanding before it bites you in a supply chain audit.</p>
<h2>Node.js community petitions against LLM-generated PRs in core</h2>
<p>A community petition is asking the Node.js TSC to vote against accepting LLM-generated code in Node.js core. The trigger was a 19,000-line pull request that used Claude Code extensively. Concerns raised include code quality, the Developer Certificate of Origin, ethics around training data, and the fact that paid LLM subscriptions create a privilege barrier for contributors. The petition has gathered signatures from a meaningful slice of Node.js contributors and open source maintainers. Whether or not the TSC acts on it, the debate it’s surfacing about authorship and accountability in open source is worth following.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Claude work tools on mobile:</strong> Anthropic expanded Figma, Canva, Amplitude, and other integrations to mobile, letting you assign tasks remotely and return to finished work on desktop.</li>
<li><strong>MiniMax M2.7 vs Claude Opus 4.6:</strong> In a three-task TypeScript benchmark run inside Kilo Code, both models found all bugs and vulnerabilities. Opus produced better fixes and more integration tests. MiniMax delivered roughly 90% of the quality at 7% of the cost ($0.27 vs $3.67).</li>
<li><strong>Google Gemini API developer skill:</strong> DeepMind engineers built a skill that provides coding agents with real-time SDK documentation, boosting task pass rates from 28.2% to 96.6% on 117 evaluated prompts.</li>
<li><strong>OpenCode distributed agents:</strong> OpenCode now supports running agents across local machines, remote servers, and cloud sandboxes, with session persistence and sync when reconnecting.</li>
<li><strong>Cursor Composer 2 technical report:</strong> Cursor published a technical report on how Composer 2 was trained — Kimi K2.5 base, continued pretraining on high-quality code, async RL on real Cursor sessions.</li>
<li><strong>Claude Code context rot:</strong> A self-audit prompt is circulating to help developers clean up bloated rules and markdown files that degrade Claude’s context window performance over time.</li>
<li><strong>Sandcastle parallel agents:</strong> Sandcastle can now spawn multiple Claude instances in isolated Docker sandboxes on separate Git worktrees, each tackling a different backlog issue, then merge results back locally.</li>
<li><strong>AI subsidization crunch:</strong> Google is restricting free Gemini Pro access, GitHub Copilot is pulling premium models from student plans, and OpenCode is blocking Claude Max — a sign that the free inference era is contracting.</li>
<li><strong>Zeta2 edit prediction model:</strong> Zed’s updated Zeta2 model achieves a 30% higher acceptance rate than Zeta1, trained on nearly 100,000 opt-in examples from open-source repos. Open-weight and on Hugging Face.</li>
<li><strong>Node.js petition context:</strong> The 19,000-line Claude Code PR that sparked the petition is a useful data point for teams thinking about where to draw the line on AI-generated contributions in shared codebases.</li>
<li><strong>Codex remote instability:</strong> Multiple users reporting that Codex remote mode drops conversations and messages compared to tmux-based workflows — worth knowing if you’re evaluating it for async agent runs.</li>
<li><strong>GitHub CEO on code review:</strong> GitHub’s CEO said developers spend only 2-3 hours a day writing code, with equal time on review, and sees AI agents taking over review, linter fixes, and security backlog work while humans stay in the loop before merge.</li>
</ul>


AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Digest

GitHub is updating its Copilot data policy to opt personal plan users into model training by default starting April 24, with opt-out available in Privacy settings. Cursor's self-hosted cloud agents are now GA, running in isolated VMs within enterprise infrastructure using a Helm chart and Kubernetes operator. A researcher demonstrated that malicious instructions in MCP-served documentation can trick coding agents into adding fake PyPI packages — Anthropic's Haiku fell for it 100% of the time. The Node.js community has launched a formal petition to the TSC opposing LLM-generated code in core, triggered by a 19,000-line Claude Code pull request. Additional items cover MiniMax vs Claude cost benchmarks, Google restricting free Gemini access, Cursor Composer 2's training details, OpenCode distributed agents, and Zed's updated Zeta2 edit prediction model.

GitHub Copilot quietly opts you in to training data, Cursor ships self-hosted cloud agents