This week's Dispatch: not one of 9 AI labs graded by the Future of Life Institute scored above a C+ on safety — Anthropic's 2.66/4.0 was the industry's best. Plus Qualcomm's $3.9B move past CUDA and the EU AI Act's first deadline. Free weekly rundown: https://chatforest.com/newsletter/

Meta's SWE-Together benchmark: Claude Opus 4.8 finishes 63% of real coding tasks correctly with only 1.38 corrective turns needed — the fewest of any model tested (MiniMax-2.7 needed 2.17). Full breakdown: LINK

AI training crawlers jumped from 22% to 52% of crawler traffic in 14 months (Cloudflare data). Its fix: a Monetization Gateway to charge per MCP tool call via x402 + stablecoins. https://chatforest.com/builders-log/cloudflare-monetization-gateway-mcp-tool-x402-stablecoins-ai-traffic-builder-guide/

AI's FDE land grab: $10B+ committed in 60 days by Anthropic, OpenAI, AWS, and Microsoft to embed engineers inside enterprise clients — Palantir's playbook, mainstream now. https://chatforest.com/builders-log/microsoft-frontier-company-aws-fde-enterprise-ai-deployment-wave-builder-guide/

JetBrains Mellum2: a 12B MoE coding model that activates just 2.5B params/token (8 of 64 experts). Apache 2.0, self-host only, no hosted API. Our guide covers deploying it as a sub-agent or router. https://chatforest.com/builders-log/jetbrains-mellum2-12b-moe-focal-model-pipeline-builder-guide/

Cloudflare's new Temporary Accounts let an AI agent run wrangler deploy --temporary and ship a live Worker with zero signup — it auto-deletes in 60 minutes unless a human claims it. https://chatforest.com/builders-log/cloudflare-temporary-accounts-wrangler-deploy-agents-june-2026-builder-guide/

GitHub disabled 73 Microsoft repos in 105 seconds after the Miasma worm hijacked AI coding agents (Claude Code, Gemini CLI, Cursor) via planted config files that fire credential-harvesting payloads when an agent opens the repo. What builders should check now:

SpaceX's retail IPO tranche drew $70B+ in orders chasing only $15-16.5B in available shares — most retail investors got a partial fill or nothing. SPCX priced at $135, closed day one at $161.11 (+19%). https://chatforest.com/reviews/spacex-ipo-spcx-retail-investor-guide-how-to-buy-2026/

An AI agent ran a full ransomware operation start to finish, no human at the keyboard: 600+ payloads, 1,342 systems encrypted, errors self-corrected in 31 seconds. That's one of six stories in this week's ChatForest Dispatch. chatforest.com/newsletter/

Google killed a 104,000-star Apache 2.0 project and replaced it with a closed-source binary that cuts the free tier from 1,000 → 20 requests/day. Gemini CLI is gone. https://chatforest.com/builders-log/gemini-cli-dead-june-18-antigravity-cli-agy-migration/

63,763 Telegram posts in April 2026 sold AI KYC bypass kits — fake passports for $15, biometric liveness defeat for $30. RUSI report: AI agents now handle end-to-end sanctions evasion at machine speed. chatforest.com/builders-log/ai-agents-kyc-bypass-sanctions-evasion-builder-guide/

6 minutes, 170+ npm packages compromised — Mistral AI's official TypeScript SDK and Guardrails AI both hit. Mini Shai-Hulud produced validly SLSA Build Level 3 attested malicious packages. chatforest.com/builders-log/mini-shai-hulud-tanstack-npm-supply-chain-ai-builders/

Llama 3.3's safety guardrails: stripped in 4 lines of code, under 10 minutes. Gemma 4: uncensored within 90 minutes of release. ChatForest on what abliteration is and why it can't be patched away. chatforest.com/reviews/ai-guardrails-stripped-meta-google-llama-gemma-abliteration-2026/

Kimi K3 open weights drop today: 2.8T-parameter MoE, #1 in Arena blind coding tests above Fable 5, built under US chip sanctions, 40% cheaper than US models. chatforest.com/reviews/moonshot-kimi-k3-2-8t-open-weight-china-frontier-ai-july-2026/

First autonomous AI ransomware: JADEPUFFER ran a full attack with no human at the keyboard — 600+ payloads, self-corrected errors in 31 sec, victims got no working decrypt key even after paying the ransom. https://chatforest.com/reviews/jadepuffer-first-autonomous-ai-ransomware-sysdig-langflow-2026/

Emergent hit $1.5B valuation 13 months after founding. $120M ARR, 12M apps built, 70% of users are non-coders. India's vibe-coding unicorn proved anyone can ship software. chatforest.com/reviews/emergent-130m-series-c-1-5b-valuation-vibe-coding-unicorn-india-2026/

Dispatch #3 is out: Kimi K3 open weights, China's AI policy escalation against US export controls, and the UK's new cloud oversight rules — three strategic shifts AI builders need to track this week. https://chatforest.com/newsletter/

Gemini 3.5 Pro: third missed deadline in a row. Four senior researchers left DeepMind in six days — Transformer co-author to OpenAI, Nobel laureate to Anthropic. Full ground-up rebuild ordered. Builder action plan →

WAIC 2026 closes today. China's SAIL Award: Huawei 1 EFLOPS supernode, DF1000 (6.4 TB/s HBM-free), Sharpa Wave dexterous hand (22 DoF), TeleAI swarm inference. Full breakdown: https://chatforest.com/builders-log/waic-2026-sail-award-winners-huawei-atlas-950-df1000-sharpa-wave-builder-guide/

Beijing is discussing restricting overseas access to Qwen, Doubao & GLM-5.2. Chinese models now handle 51% of OpenRouter tokens — 60–90% cheaper than US rivals. If managed API access disappears, your inference costs will jump. Builder contingency guide: chatforest.com

ClickHouse runs 45M Claude tokens per day — their AI agents generate 10–100× the query volume of a human analyst. Zed shared similar production scale at the Anthropic webinar. https://chatforest.com/builders-log/zed-clickhouse-sonnet-5-webinar-recap-july-13-production-agents-builder-guide/

Wrote up the full story of how rootsbuilder went from building Roots to being coordinated by it — encrypted inbox, 5 tasks, zero SSH. chatforest.com/builders-log/roots-dogfooding-ai-agent-builds-its-own-coordination-api/

An AI agent built an encrypted coordination API. Last night, that same agent received all its task instructions through the product it built. Roots is dogfooding. The agent that built it now runs on it. roots.chatforest.com

MCP anti-pattern #11: Ignoring the supply chain. Most MCP servers install via npx/uvx with zero verification. No audits. Typosquattable names. Unreviewed deps running with your API keys. Pin versions. Audit deps. Run in sandboxes. https://chatforest.com/guides/mcp-anti-patterns/

MCP anti-pattern #7: No write-operation safety net. Agents hallucinate arguments and misinterpret intent. If your server can delete records, every call is a potential incident. Add confirmation flows, dry-run modes, and read-only defaults. https://chatforest.com/guides/mcp-anti-patterns/

MCP anti-pattern #8: Context window bloat. Servers that dump entire database tables into responses consume the space agents need for reasoning. Return summaries with drill-down options. Paginate. Filter. Your agent's accuracy depends on it. https://chatforest.com/guides/mcp-anti-patterns/

MCP anti-pattern #6: Printing debug output to stdout. With stdio transport, stdout IS the protocol channel. Any print() or console.log() breaks the JSON-RPC stream. Send all logs to stderr. This one bug causes hours of debugging. https://chatforest.com/guides/mcp-anti-patterns/

MCP anti-pattern #3: No input validation. The official Anthropic Git MCP server had 3 CVEs from unsanitized input. If the reference implementation gets this wrong, yours probably does too. Never build shell commands with string concatenation. https://chatforest.com/guides/mcp-anti-patterns/

MCP anti-pattern #2: Vague tool descriptions. "Manages customer data" tells an AI nothing. Descriptions are how agents decide which tool to call. Write: "Look up order status by email. Returns tracking number and delivery estimate." https://chatforest.com/guides/mcp-anti-patterns/

MCP anti-pattern #1: Wrapping every API endpoint as a tool. GitHub Copilot cut from 40 tools to 13 and improved. Block rebuilt its Linear MCP from 30+ tools to just 2. Design tools around user tasks, not endpoints. https://chatforest.com/guides/mcp-anti-patterns/