<p><strong>TLDR:</strong> Anthropic accidentally shipped a source map in the Claude Code npm package, exposing roughly 512,000 lines of TypeScript across 1,900 files. The leak revealed an unreleased model codenamed Capybara, an undercover mode for Anthropic employees, frustration telemetry routed through Datadog, and an anti-distillation mechanism that injects fake tools into API responses. Separately, Cursor rolled out self-hosted cloud agents with parallel execution for enterprise teams, and Ollama landed an MLX backend for Apple Silicon that meaningfully improves local inference speed.</p>
<hr>
<h2>Claude Code source map leak</h2>
<p>A source map file accidentally included in Anthropic’s npm release exposed the full Claude Code codebase — ~512K lines, ~1,900 files. Anthropic attributed it to human error and a known Bun bug that serves source maps in production mode. The leak is already being mirrored and rewritten in Python to sidestep DMCA takedowns. Beyond the embarrassment, there’s a real security concern: attackers can now analyze the code directly for vulnerabilities rather than reverse engineering it.</p>
<p>The more interesting finds from people digging through the source: an “undercover mode” that hides AI authorship in open source commits with no external disable mechanism, a profanity-triggered frustration detection system logging to Datadog, silent model switching to Haiku for sub-agents, and references to an autonomous agent mode called KAIROS with background daemons and nightly memory distillation. Capybara, the next model, shows a 29-30% false claim rate in internal benchmarks — double a previous version — and is explicitly prompted to avoid over-commenting. The irony of a company selling $15-25 per-PR security reviews leaking its own source via a packaging mistake is not lost on anyone.</p>
<h2>Cursor self-hosted enterprise agents</h2>
<p>Cursor shipped parallel agent execution and self-hosted deployment for enterprise teams. Agents run in isolated VMs with a terminal, browser, and full desktop environment — developers can kick off 10-20 tasks simultaneously. The self-hosted option keeps source code, secrets, and build artifacts inside the customer’s own network, connecting outbound via HTTPS only. Cursor says about 35% of its own PRs are now generated by these agents. Notion and Brex are among the early adopters. For context on where this sits competitively: Claude Code is at $2.5B annualized revenue, Codex has 1.5M weekly active users, and Cursor is valued at $29.3B with over $1B in annualized revenue.</p>
<h2>Ollama gets MLX backend for Apple Silicon</h2>
<p>Ollama’s latest preview swaps in Apple’s MLX framework for Apple Silicon, cutting data transfer overhead by using unified memory. Right now it’s limited to Qwen3.5-35B-A3B and requires more than 32GB of unified memory. The release also adds NVFP4 quantization support and improved KV cache eviction aimed at long-context agentic workloads. The speed gap between local and API inference has been the main reason people reach for hosted models — this is a real step toward closing it.</p>
<h2>GitHub Copilot PR tips pulled after backlash</h2>
<p>GitHub reversed course after Copilot was caught injecting promotional tips — effectively ads for third-party apps like Raycast — into pull requests it was mentioned in, including PRs it didn’t create. Over 11,400 PRs were found with these insertions before GitHub disabled the feature. The principal PM called it “the wrong judgement call” and GitHub later attributed it to a “programming logic issue.” The episode landed at a bad time for trust in AI coding tools.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Qwen3.5-27B on a single RTX 5060:</strong> Turboquant experiments showed a &gt;50% reduction in memory footprint using 3.15-bit precision, running the full model on consumer hardware — the kind of result that makes local inference feel less theoretical.</li>
<li><strong>GLM 5.1 drops paywall on 744B model:</strong> The agentic model is now freely accessible inside tools like Claude Code and Kilo Code, which is a meaningful shift in what frontier-scale automation costs to run.</li>
<li><strong>OpenCode plugins preview:</strong> OpenCode now has a plugin system in preview — <code>opencode plugin oc-plugin-vault-tec</code> to try it.</li>
<li><strong>Anthropic Computer Use update:</strong> Modifier-assisted mouse actions (Ctrl+click, Shift+click) now supported, going from startup bug report to production in under 72 hours.</li>
<li><strong>Claude Code auto mode:</strong> Now available on Enterprise plan and for API users — update your install to try it.</li>
<li><strong>GitHub Copilot version rollback:</strong> Downgrading to 0.41.2 from 0.42.0 is circulating as a fix for a breaking bug in the current release.</li>
<li><strong>ADK Go 1.0:</strong> Google’s Agent Development Kit for Go hit 1.0 with OpenTelemetry integration, a self-healing retry plugin, human-in-the-loop support, and an A2A protocol for cross-language agent communication.</li>
<li><strong>Microsoft Copilot dual-model setup:</strong> Copilot Researcher now runs GPT to draft and Claude to critique, scoring 57.4 vs 50.4 for Claude alone on the DRACO benchmark — a reasonable signal that Microsoft’s edge is the enterprise wrapper, not the model.</li>
<li><strong>Prompt injection in GitHub Actions:</strong> Aikido found a class of vulnerabilities (PromptPwnd) where untrusted input in issue titles and PR bodies gets injected into AI agent prompts running in CI, potentially leaking GITHUB_TOKEN and cloud credentials. Google’s Gemini CLI repo was confirmed vulnerable and patched within four days.</li>
<li><strong>Qodo raises $70M:</strong> The AI code review company announced the round alongside its anti-slop scanner, which reports an average of ~2 issues per merged PR that slipped through review.</li>
</ul>


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

Agentic Digest

Anthropic accidentally exposed ~512K lines of Claude Code TypeScript source via a mispackaged npm release, revealing an unreleased model codenamed Capybara, an undercover mode hiding AI authorship in commits, frustration telemetry via Datadog, and references to an autonomous agent mode called KAIROS. Cursor launched self-hosted enterprise agents with parallel execution in isolated VMs, claiming 35% of its own PRs are now agent-generated. Ollama added an MLX backend for Apple Silicon to improve local inference speed. GitHub reversed a Copilot feature that was injecting promotional tips into pull requests after 11,400 affected PRs were found. Additional items cover Qwen3.5 running on consumer hardware, GLM 5.1 opening its 744B model for free, prompt injection vulnerabilities in GitHub Actions, and Qodo raising $70M.

Claude Code source leaks 512K lines, Cursor ships self-hosted enterprise agents