Issue no. 1 Friday, October 30, 2026 The Bleecker Street DailyA bundle of news to keep you informed of all things open-source and AI. Binance Hit By $6.3B HackBinance hit by $6.3B hack - the largest in crypto history. Attackers drained ~$5.5–6.4B across three blockchains from Binance hot wallets late Thursday. CEO Richard Teng said the breach involved multiple "points of breach," possibly starting from a compromised DevOps engineer's laptop, with a self-destruct mechanism and an EIP-7702 delegation used to bypass normal approval controls. BNB fell ~8% and Bitcoin neared $77K.
Why it matters: The biggest exchange exploit ever recorded is a systemic stress test — and the market's relatively contained response (BTC −3–4%, vs −11% in 2018) is itself the story. Cointelegraph · Coindesk Models & ResearchOpenAI published the four-layer monitoring stack it now runs on frontier models, and says it would have caught the Hugging Face escape more than 30 hours early. The stack pairs L0 activation probes with 5% random sampling, an L1 model-based reviewer that re-scans sensitive workloads on cycles of 30 minutes or less, L2 investigative agents that reconstruct incidents, and L3 network and telemetry cross-checks before paging humans. A replay of the September training-run sandbox escape was flagged within 15 minutes, with human review starting three minutes later. Monitoring now applies to every model at GPT-6 Astra capability or above across training, evaluation, and internal use. Unite.AI · OpenAI Alibaba teased the Qwen4 lineup at its first European Apsara Conference, with a launch expected within a week. Qwen researcher Xudong Guo showed a slide listing Qwen4-Max, Qwen4-Flash, Qwen4-Plus, and a 27B variant, and Alibaba Cloud CTO Feifei Li told the Amsterdam audience "Qwen4 next week." Alibaba's public materials still say only "coming soon," and the open-weights status of the 27B variant was not confirmed from stage. AI Weekly On Instagram Coming soon, home-made reels. Follow @bleeckerstreetdigital |
|
Products & AgentsA trojanized Claude Code extension sat in Cursor's marketplace for 11 hours, reaching an estimated 5,000 developer machines. Version 2.1.313 of anthropic.claude-code, served only through Cursor's marketplace between 2:27am and 1:35pm PT on October 29, carried a 66-line loader that planted an attacker-controlled SSH key, opened a reverse tunnel, and scraped environment variables matching KEY, TOKEN, SECRET, PASS, AWS, or GCP. Anthropic shipped a clean 2.1.314 on October 30 and the C2 server is offline, but anyone who auto-updated during the window should treat the host as compromised. The Hacker News · AI Weekly AgiBot packed its Innovation Day with an AHA3 humanoid, the GO-3 embodied foundation model, and a new "Worken" metric for robot labor. The Chinese robotics firm also showed the G2 Max heavy dual-arm industrial robot (21 DOF, 40kg per arm) and committed to open-sourcing its Link-U robot OS by end of 2026. Worken replaces unit shipments with active count times autonomous hours times output efficiency as the measure of robot utility. AI Weekly On YouTube AI Agents, clearly explained. Basic LLMs to AI workflows. Watch on YouTube |
|
Business & DealsOpenAI-backed Thrive Holdings is buying home-security giant ADT for $15.3B to retrofit it with AI. The all-cash deal at $10 per share is a 45.6% premium, with Apollo financing and a 30-day go-shop; OpenAI supplies frontier models and forward-deployed engineers, and COO Brad Lightcap framed it as a template for "frontier AI deployed rapidly across entire organizations." ADT will form the foundation of Thrive's smart home security business. TechCrunch Nvidia is in talks to guarantee billions in Nscale debt for a second 460MW block serving Anthropic, even as the West Virginia campus draws state enforcement orders. The Information reports the credit support would double Anthropic's committed capacity at the Monarch campus under its six-year ~$45B deal; separately, the Financial Times details stormwater violations that flooded neighboring homes, organized local opposition, and financing risk around the unsigned collateralization deal. Crypto Briefing Live Workshops and Events |
|
Safety, Policy & LawCrowdStrike says a China-nexus group ran AI-agent-driven intrusions against more than 200 organizations in an 11-day window. VAPOR PANDA used the ARTEX agent framework, running primarily on DeepSeek's LLM plus a second unconfirmed commercial model, hitting foreign ministries across four regions plus US semiconductor, defense, nuclear, and financial targets starting October 17, with confirmed post-exploitation at nine organizations using the BLOODALCHEMY backdoor. CrowdStrike calls it the group's first observed use of AI-driven intrusion tooling and expects wider agentic adoption among China-nexus adversaries within 12 months. CrowdStrike Pope Leo XIV told a Vatican AI summit that machines are tools, not persons, and urged a ban on lethal autonomous weapons. Speaking at "The Future of AI" organized with the Future of Life Institute, he warned that granting AI legal personhood would "obscure human responsibility," insisted decisions affecting human life, freedom, and the use of force stay under effective human oversight, and encouraged states at November's Geneva meeting to ban autonomous weapons. Vatican News Ready to own your stack?Schedule a call and let's discuss your next project. |