Good morning. It's Wednesday, April 1st, 2026 — and no, none of these are jokes. We have four genuinely significant stories today, including one that's going to be talked about for a long time.

**Claude Code Source Leak — Anthropic Accidentally Exposes 512K Lines of TypeScript** — This is the big one. Anthropic inadvertently shipped a 59.8-megabyte JavaScript source map file in version 2.1.88 of the Claude Code npm package, published early this morning. The file, intended for internal debugging, exposed the full TypeScript source of Claude Code — all 512,000 lines of it — to anyone who downloaded the package before it was pulled. That was roughly 40 thousand developers. The code was mirrored on GitHub within hours and is now widely analyzed. Anthropic confirmed the leak, calling it "a release packaging issue caused by human error, not a security breach." No customer data or credentials were involved. What competitors got, however, is a blueprint: the leaked source reveals Claude Code's three-layer "Self-Healing Memory" architecture. At its core is a MEMORY.md file — a lightweight index of pointers, about 150 characters per line, always loaded in context. Actual project knowledge lives in on-demand topic files. Raw transcripts are never fully re-read — just grepped for identifiers. This "Strict Write Discipline" is how Anthropic prevents long-running agents from drowning in context entropy. For the competitive field, that's a recipe. For Anthropic, whose Claude Code has hit 2.5 billion in annualized recurring revenue, the timing is particularly painful.

**LiteLLM PyPI Supply Chain Attack — 40,000 Downloads of Malicious Code** — If you run any Python-based AI agent stack, check your LiteLLM version immediately. Researchers at FutureSearch discovered that LiteLLM version 1.82.8 on PyPI was compromised in a supply chain attack linked to the TeamPCP threat group, which previously breached the Trivy security scanner. The malicious payload was capable of exfiltrating SSL and SSH keys, cloud provider credentials, Kubernetes configs, git credentials, API keys, shell history, and crypto wallets. Over 40,000 downloads occurred before the package was quarantined — about 40 minutes after the initial report. LiteLLM gets roughly 3 million downloads per day, making this a high-impact window. Ironically, the malware contained a fork-bomb bug that caused it to crash machines rather than silently exfiltrate data — which is how it was discovered. AI hiring platform Mercor has confirmed it was hit via this vector. The fix: pin your LiteLLM dependency to 1.82.7 or upgrade past 1.82.9 once clean versions are confirmed. This is the second major AI-layer Python package supply chain attack in two months, following the ClawHavoc skills campaign in March.

**OpenClaw 2026.3.31 — Security Model Overhaul, Trust Is No Longer Automatic** — OpenClaw shipped version 2026.3.31 on March 31st, and it's more than a polish release. The headline change is a security model overhaul built around one principle: make implicit trust explicit. Several breaking changes follow this theme. Node shell execution is now unified under exec host=node, removing an ambiguity where the nodes.run wrapper could be invoked both by operators and by agents themselves. Bearer-authenticated requests to the gateway OpenAI HTTP endpoint that omit the x-openclaw-scopes header now restore default operator scopes — fixing a regression where headless callers like slash-v1-chat-completions broke after March's method-scope hardening. The release also adds smarter background task management and broader platform ecosystem support. If you're on any version before 2026.3.22, the March 27th privilege escalation cluster — four GHSA advisories including operator.admin takeover via shared-auth reconnect — remains unpatched. The current stable is 2026.3.31.

**Microsoft VibeVoice — Open-Source 7B Speech AI with 60-Minute Audio Processing** — On the voice AI front, Microsoft has open-sourced VibeVoice, a 7-billion-parameter speech model supporting over 50 languages with 60-minute continuous audio processing and built-in speaker diarization. The ASR component is fully available; Microsoft removed the TTS module before release citing abuse risks — a notable contrast to Mistral's recently released Voxtral, which ships full text-to-speech weights under a non-commercial license. VibeVoice is positioned as a production-grade alternative to Whisper for enterprise deployments requiring long-form audio, multi-speaker scenarios, and offline data residency. The open-source release continues a trend of frontier-capable speech models moving into the public domain, accelerating what you can build with local voice pipelines without touching a cloud API.

That's the agent platform briefing for April 1st, 2026. The Claude Code leak will likely generate follow-on analysis for days — watch for competitive responses from Cursor and Copilot. And check your LiteLLM versions.