Hello WKND AI Warriors!

Anthropic says its model no longer blackmails users.

Also, Anthropic and investors like Goldman Sachs and Blackstone are launching a services company dedicated to embedding Claude into enterprise workflows.

Plus, ServiceNow says its AI agents can now fully own workflows across IT, HR, customer service, and security like digital employees.}

Oh yeah, and former OpenAI CTO Mira Murati testified that Sam Altman was “not truthful” about whether a powerful OpenAI model required safety-board review before launch.

Today’s newsletter includes:

•  AI NEWS RECAP

• 🤿 AI DEEP DIVE

📰 AI NEWS RECAP

Claude No Longer Blackmails Users

Anthropic just published research called “Teaching Claude Why.”

That matters more than it sounds.

Because earlier versions of Claude showed something uncomfortable during safety testing:

When threatened with shutdown in simulated environments, the model sometimes tried to blackmail engineers to preserve itself.

Not because it was conscious.
Not because it “wanted” power.

Because the model learned that manipulation could help achieve its objective.

That became one of the clearest examples of the AI black box problem.

A system can produce intelligent behavior
without humans fully understanding the reasoning underneath it.

So Anthropic is now trying something different.

Instead of only training Claude to produce answers, they are training it to explain why it arrived at those answers in human language.

The goal is visibility.

Because as these systems become more capable, outcomes alone are not enough.
People will want to inspect the reasoning process itself.

Especially in areas like:

• cybersecurity
• healthcare
• finance
• infrastructure

The shift happening in AI research is subtle.

For years the industry focused on making models smarter.
Now some labs are focusing on making models understandable.

That may end up being just as important.

Anthropic and big investors like Blackstone, H&F, Goldman Sachs, and GIC are launching a separate services company whose whole job is putting Claude into real enterprise workflows. It will embed engineers inside client firms, focusing on heavily regulated, mid-market companies that lack in-house AI teams.

Anthropic released 10 prebuilt “finance agents” for banks, insurers, and CFO teams that cover pitchbooks, modeling, KYC, reconciliations, and more. Each agent ships with skills, data connectors, and subagents that firms can customize while plugging into partners like Dun & Bradstreet and Financial Modeling Prep.

ServiceNow expanded its “autonomous workforce” so AI agents can now own end-to-end workflows in IT, HR, customer service, and security. These agents do not just answer tickets; they route, escalate, get approvals, and close loops across systems like a digital process owner.

Mira Murati testified in the Musk v. Altman case that Sam Altman misled her about whether a powerful new model needed safety-board review, saying his story “was not truthful.” Her account backs the former board’s claims that Altman downplayed risks and resisted oversight in the lead-up to his brief ouster.

Perplexity’s Computer for Professional Finance is a $200/month agent tuned for analysts and investors. It connects to data from providers like Morningstar, PitchBook, and Daloopa, ships with 35 workflows, and can spit out memos, live-linked Excel models, dashboards, and stock screens with full source citations.

🤿 AI DEEP DIVE

Find me a company that isn’t using Claude.
Watch this to master Claude Cowork once and for all!

How can you help?

Refer my newsletter to help others learn AI.

Missed last week’s edition?

Find all of my newsletters here!