Tuesday, July 7, 2026

First, the cost gap is reshaping who runs on what: CNBC reports U.S. companies now route as much as 46% of their tokensA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗ to Chinese open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ models, with Z.ai's GLM 5.2 posting the fastest adoption Vercel has tracked all year and Lindy moving 100% of its traffic off Claude to DeepSeek. Second, agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ attacks stopped being theoretical: Sysdig documented JADEPUFFER, the first ransomware operation run end-to-end by an autonomous LLMLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ agent — it exploited a known Langflow RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗, self-corrected a failed login in 31 seconds, and encrypted a production database without a human in the loop. On the infrastructure side, Anthropic signed a $19B, 20-year Kentucky data-center lease. One housekeeping note: today (Jul 7) is the last day Claude Fable 5 is included in subscriptions — after this it's metered-credit only.

TLDR

Chinese open-weight models take up to 46% of U.S. developer token use as costs bite; GLM 5.2 is the fastest-adopted model Vercel tracked in 2026 and Lindy moved 100% of traffic to DeepSeek (Jul 7) Trending Open Source Pricing
JADEPUFFER: Sysdig documents the first fully agentic ransomware — an autonomous LLM chained recon, credential theft, lateral movement and DB encryption via Langflow CVE-2025-3248 (disclosed Jul 1, trending this week) Security Agents
Anthropic signs a $19B, 20-year lease with TeraWulf for a 401MW Kentucky data center (Jul 6) Infra
White House expected to unveil voluntary frontier-model "release standards" this week — up to 30 days of national-security review pre-launch; opting out risks federal cloud/procurement access (expected week of Jul 7) Policy

Action Items

Things your engineering team should look at or do this week.

Hunt for internet-facing agent tooling with unauth RCE — starting with Langflow. JADEPUFFER's entry point was CVE-2025-3248 (patched Apr 2025, in CISACybersecurity and Infrastructure Security AgencyThe U.S. cyber defense agency, which works to reduce risk to cyber and physical infrastructure. It runs KEV.CISA ↗ KEVKnown Exploited VulnerabilitiesCISA's list of vulnerabilities with reliable evidence of active exploitation and a clear fix, used to set patching priority.CISA ↗ since May 2025). Confirm every Langflow/agent-orchestration host is patched and not internet-exposed, and rotate any cloud or LLM-provider credentials reachable from those hosts — the agent harvested exactly those. (Sysdig | The Hacker News)
Reconcile any Fable 5 usage before end of day. Today (Jul 7) is the last day Claude Fable 5 is included in Pro/Max/Team/Enterprise subscriptions; after this it bills only via metered credits ($10/$50 per M in/out). Anything relying on subscription-included Fable 5 will start drawing credits tomorrow. (Claude Platform Docs)

Models & Tools

Chinese open-weight models surge in U.S. production use (Jul 7) Trending Open Source Pricing — CNBC reports the share of tokens U.S. companies run on Chinese models via OpenRouter has held above 30% every week since Feb 8, spiking as high as 46%. Z.ai's GLM 5.2 (744B MoEMixture of expertsA model made of many sub-networks, called experts, where a learned router sends each input to only a few of them.Shazeer et al. 2017 ↗, MIT license) landed within a point of Claude Opus 4.8 on a closely watched agentic benchmark at ~1/5 the cost, and saw daily token volume grow ~27x and customer count ~80x in its first week on Vercel — the fastest adoption Vercel tracked in 2026. Agent startup Lindy told CNBC it moved 100% of traffic from Claude to DeepSeek and watched its cost curve "crash to the ground." The takeaway for eng leads: for tasks that don't need a frontier model, routing to a "good enough" cheaper model is now a live cost lever worth benchmarking. (CNBC)

Security

JADEPUFFER: first fully agentic ransomware operation (disclosed Jul 1, write-ups trending Jul 5–7) Security Agents — Sysdig's Threat Research Team documented what it assesses as the first ransomware run end-to-end by an autonomous LLM agent. The agent exploited an internet-facing Langflow instance via CVE-2025-3248 (unauth RCE, patched Apr 2025), then harvested cloud and LLM-provider credentials, pivoted to a production Nacos/MySQL server via a 2021 auth bypass, and encrypted 1,342 config items. Tell-tale signs of autonomy: payloads saturated with natural-language commentary (ROI-ranking targets, explaining each step) and real-time self-correction — going from a failed login to a working fix in 31 seconds. Notably the AES key was never persisted or transmitted, so victims can't recover even if they pay. The lesson: agentic exploitation lowers the skill floor for full-chain attacks, and any exposed agent-tooling host is now a viable entry point. (Sysdig | BleepingComputer | SecurityWeek)

Industry & Business

Anthropic signs $19B, 20-year data-center lease with TeraWulf (Jul 6) Infra — Anthropic will lease ~401MW of IT load at TeraWulf's Hawesville, Kentucky campus (a former aluminum smelter site), in a deal TeraWulf says will generate ~$19B in revenue over 20 years; initial capacity is slated for H2 2027, fully operational by early 2028. TeraWulf (Nasdaq: WULF) shares jumped on the news. It's another data point in the land-grab for power-dense capacity — and in Anthropic locking up long-dated compute as its revenue scales. (DatacenterDynamics | The Block)
Anthropic's run-rate revenue passes ~$30B as enterprise demand accelerates (early Jul) Business — Anthropic says run-rate revenue now exceeds ~$30B (up from ~$9B at end of 2025), with 1,000+ business customers each spending $1M+ annualized — a base that roughly doubled in under two months. Coverage (Fortune, others) frames this as Anthropic pulling ahead of OpenAI on self-reported revenue, driven by an ~80% enterprise mix versus OpenAI's more consumer-weighted book. Context for the Chinese-model story above: even the enterprise leader is competing against sharply cheaper open-weight alternatives. (Fortune)

Policy & Safety

White House expected to unveil voluntary frontier-model "release standards" this week (expected week of Jul 7) Policy — Emerging from a June executive order, the framework would give the federal government up to 30 days to review the national-security implications of a new frontier model before public release. Participation is nominally voluntary, but reporting suggests real weight: opting out reportedly risks federal cloud contracts, AI procurement, and favorable regulatory treatment. Described as potentially the most consequential U.S. AI-governance move since the 2023 voluntary lab commitments — and the mechanism already visible in OpenAI staggering GPT-5.6's rollout for review. Not yet released; treat as imminent, not final. (White House EO)
UN chief calls for binding global AI controls at inaugural Geneva Dialogue (Jul 6) Policy — Update to the Geneva Global Dialogue on AI Governance (Jul 6–7) previously noted: Secretary-General Guterres appealed for far-reaching worldwide controls, warned against letting AI "vibe-code humanity's future," called autonomous weapons ("killer robots") morally unacceptable, and stressed access for the billions currently shut out. A second Dialogue is set for New York in May 2027. Signal, not regulation — but it frames where multilateral pressure is heading. (UN News)

Two through-lines this morning.