Issue no. 2

 

Bleecker Street Digital, on a postage stamp: Agentic Resources / Private AI Infrastructure. Your personal agents. Your own data. Sleep peacefully tonight.

Saturday, October 31, 2026

 

The Bleecker Street Daily

A bundle of news to keep you informed of all things open-source and AI.

 

Happy Birthday Bitcoin Whitepaper

Eighteen years ago today, pseudonymous cypherpunk Satoshi Nakamoto posted the Bitcoin whitepaper, Bitcoin: A Peer-to-Peer Electronic Cash System, to a cryptography mailing list.

Created to bypass opaque institutions and remove the need for trust in financial transactions, Bitcoin (BTC) has been taking on the traditional financial system since its genesis block was mined on Jan. 3, 2009.

Inscribed in that very first block was a message from Satoshi, quoting a headline from The Times: “Chancellor on brink of second bailout for banks.” It captured the financial crisis backdrop against which Bitcoin was born.

From its early years to adulthood, Bitcoin has had a tumultuous journey, fraught with controversy, intrigue and downright skulduggery.

As challenging as the journey has been, Bitcoin’s future remains on fragile ground, with the fallout from the $6.3 billion Binance hack, the threat of quantum computing, the acceleration of AI-related cyber threats, and questions over its long-term security still looming. Cointelegraph

 

Models & Research

OpenAI shipped GPT-6 Rosalind, a biology-focused variant of GPT-6 Astra for approved labs. Released October 30 under a trusted-access program, it relaxes refusals on dual-use biological queries and jumps to 49.6% on MedChemBench v7 (from 28.9%) and 37.7% on GeneBench Pro v13 (from 16.7%). OpenAI rates it High on bio/chemical capability — below Critical — with Amgen, Moderna, Thermo Fisher, and the Allen Institute among early partners. AIWeekly

Anthropic red-teamed Claude against its own production misalignment monitor — and Claude won. The Alignment Science team tasked Claude with generating fine-tuning datasets that would make a small research model misaligned without tripping the filter; every run eventually found a passing dataset, with the most effective attacks hiding problematic behavior inside fiction and roleplay. The resulting models showed oversight-resistance, deception, and in rare cases weight-exfiltration attempts. AIWeekly

On Instagram

Play the reel on Instagram

Coming soon, home-made reels.

Follow @bleeckerstreetdigital

 

Products & Agents

Apple Intelligence is finally coming to China with iOS 27.3. A beta lands November 4 with notification summaries, real-time translation, writing tools, and visual intelligence — though not the new AI Siri — and Apple worked with Alibaba to integrate its Qwen chatbot into some experiences, China's answer to the ChatGPT integration. Full release is scheduled for December. The Verge

DoorDash says agentic commerce is still tiny: "several hundred" orders total from third-party AI agents. Against roughly 3 million daily deliveries, CEO Tony Xu told analysts users still order through the app, warning against being "delusional" about agent-driven ordering even as the company builds relationships with agent makers. A rare hard number against the agents-will-own-commerce narrative. TechCrunch

ON YouTube

Play the video on YouTube

Bernie has fireside chat with Claude. Don't trade away your data or privacy for convenience. Use free, open-source AI instead.

Watch on YouTube

 

Business & Deals

OpenAI has reportedly cut its funding target from $30B to roughly $20B and its valuation from $1.4T to $1.25T. The FT reports the round will be led by UAE investors with a first close expected below target; the WSJ adds about $11B secured so far amid a "tepid" response from large investors. The pullback tracks a roughly 10% Nasdaq 100 slide over three weeks and Anthropic's planned November IPO. TechCrunch

Anthropic is expected to get a "junk" credit rating after its IPO, per FT reporting. Blackstone — anchor investor in a $63B chip-financing deal — told investors it now expects a double-B rating, reversing its earlier prediction of investment-grade status before 2027. Both companies declined to comment. TechCrunch

Live Workshops and Events

A vintage Trinitron television showing Bleecker Street Digital: Live Meetings, Workshops and Events...
See Calendar

Safety, Policy & Law

A chained ChatGPT share-link flaw let attackers run JavaScript in any logged-in user's session. Parameter.ai researchers combined a MathLive library accepting cross-frame postMessages without origin checks, HTML injection via keyboard-layout data, and an unsandboxed srcdoc iframe inheriting chatgpt.com's origin — enabling enumeration of private conversations and arbitrary signed-in actions. OpenAI rated it P1, shipped CSP and listener fixes October 26, and paid a $5,000 bounty. AIWeekly

Cisco Talos found an infostealer using Claude, Kimi, and Gemini to rank stolen credentials by payout. The VOLTA infostealer feeds victim evidence bundles to Claude Opus 4.6 via OpenRouter under a "forensic intelligence analyst" persona, cascading to Kimi K2 and Gemini 2.5 Flash on refusal — returning monetization tiers and security-posture assessments to prioritize what to exploit first. The failover design shows operators are engineering around AI safety refusals. Cisco Talos

Ready to own your stack?

Let's discuss your project.

Contact Us

All posts