Econify.AI Newsletter, October 23, 2026

AI attacks hit OpenCode and PHP

Attackers broke into OpenCode Console, where developers store the API keys that bill their AI usage. OpenCode called it a distillation attack: the stolen keys pulled huge volumes of answers from OpenAI, Anthropic and Google models, likely to train a rival model. Days later, researchers disclosed a flaw in PHP, the software behind much of the web, that reportedly affects up to 70% of websites, and the NSA advised taking exposed systems offline until patched. Hosting platform Railway said attackers may have accessed and decrypted some customer secrets.

Econify's take: The speed AI is finding and exploiting security flaws is far outpacing patching. Leadership needs to make keeping on top of these vulnerabilities a priority.

The big AI labs get credible rivals

Microsoft's AI chief Mustafa Suleyman unveiled MAI Base 2, the foundation for a service called Microsoft Frontier Tuning: a company trains Microsoft's model further on its own data and ends up with a private version only it controls. Meta released Muse Spark Open, a free model anyone can download and run. Researcher Nathan Lambert called it the first US model of its kind that comes close to the best at multistep work, an area Chinese models had led. Both answer a problem spreading AI use creates: sending every task to the most expensive model costs too much and hands one vendor your data.

Econify's take: Our read is that most mid-market firms will get this through Microsoft, the vendor already holding their data. Fine-tuning only makes sense for niche cases where a generalized model isn't good enough.

Claude Duet takes on admin work

Anthropic opened a waitlisted preview of Claude Duet for US subscribers on certain plans. It sends a morning brief from your connected apps and works in the background on tasks like chasing unpaid invoices in QuickBooks or Stripe, or preparing payroll in Gusto. Before it sends, pays, signs or deletes anything, it waits for your approval. It never directly sees your passwords, buys with a virtual card so your real card number stays hidden, and refuses to send card or government ID numbers. It checks your apps every few hours to save on plan usage limits, and Anthropic warns it can miss what arrives in between.

Econify's take: We'd join the waitlist now and start narrowly once access arrives. It's a crowded field, so if you aren't all in on Anthropic, wait a little to see what real-world reviews discover.

Meta's Muse exposed business documents

Researchers at security firm Claroc showed that the links Muse users create to share conversations could be found with a single search. They found 6,800 of them and opened a few hundred, which included a small-business owner's loan application, a company's internal revenue review and a workplace's staff vacation calendar. Mark Zuckerberg had said Muse was "built from the ground up for privacy and security." Meta paused the sharing feature after Ars Technica asked about the findings.

Econify's take: We see too many companies not officially using AI at a corporate level and instead leaving individual employees to use AI as they like. That will turn into a real risk soon if you aren't careful.

A finance firm's AI biology experiment

Citrini Research, which publishes investment analysis, believed life sciences stocks were undervalued but couldn't get a straight answer from experts on whether AI really speeds up drug discovery. So it ran the test itself, with Anthropic. The team adapted a 35-page prompt Anthropic had already tested in an earlier study. Claude and 300 AI helpers then designed 318,000 candidate proteins aimed at scorpion venom. Independent lab testing found 17 that bound to four toxins, for under a couple hundred thousand dollars, though none has yet been shown to neutralize venom. Citrini says it remains a financial research firm, and commentators expect more side projects like this one that sharpen a firm's core work.

Econify's take: No real anti-toxin was created; the motivation was to prove their holdings were undervalued. Still worth noting another use case for AI in the area of medicine.

Other news

Policy and public money

• DHS ordered federal civilian agencies to remove DeepSeek products and stop buying them, citing national security risks such as data sent to servers in China.
• After the White House AI task force's first meeting, Emil Michael said the government can use export controls to rein in AI labs.
• Anthropic is committing 5% of its shares, with more possible at the board's discretion as its value rises, to a public benefit fund that can't use them for lobbying, and Elon Musk pledged one SpaceX share to every US child.

Money and infrastructure

• CoreWeave's bonds fell to 80 cents on the dollar, and Microsoft pulled out of Nscale's Essex data center after grid power delays.
• A leaked draft prospectus shows Anthropic lost about $80 billion in the first half, mostly from a noncash accounting charge, as it prepares to go public.

Products and research

• A developer used Claude Opus to build free clones of Adobe's software suite, starting with Photoshop.
• Cognition now raises Devin usage limits for engineers who use their token budgets efficiently, and added Devin Coach to flag habits that waste usage.
• Goodfire researcher Nick Cammarata argues most neurons in large models are readable, though Google DeepMind's Neel Nanda called it interesting but not a significant update to the big picture.