Agents Got Email Addresses. Their Builders Are Quitting.

OpenAI says its agents solved a $1M math problem this week. The mathematician who was closing in on it first says OpenAI used his own unpublished work to beat him to it.

That's the thread worth pulling. This week the signal was mostly about who gets to act, and who's actually watching them do it. Instinct gave its agent a real email address so it can sign up for services and pay for them without you in the loop. n8n shipped an assistant that runs a workflow, reads why it broke, and fixes it — the first version built to skip the "paste the error back into a chatbot" step. OpenAI put a live dashboard behind a single prompt, no analyst required. Genuine capability jumps, all three. But the same week, a Zapier survey of 518 governance-minded execs found over a third of "let AI act unsupervised" deployments already caused real damage, hackers found a way to quietly drain Claude subscribers' token budgets with no way for victims to prove it wasn't them, and a researcher who spent three years inside both OpenAI and Anthropic quit calling the whole race reckless. Authority is shipping faster than anyone's checking it.

Models + launches

OpenAI says an unreleased model fielded roughly 10,000 AI agents for 88 hours — about $22.5 million in compute — and produced a proof for the Navier-Stokes existence-and-smoothness problem, one of math's seven Millennium Prize problems [1]. NYU mathematician Tristan Buckmaster says that's not the full story: he'd been quietly working the same rare approach with Anthropic's Levent Alpöge, told an OpenAI mathematician about it, and three days later OpenAI's Sébastien Bubeck said an internal model already had a proof — then allegedly pushed Buckmaster to drop Alpöge's name over his Anthropic affiliation, asking "why would you ruin your career?" when he balked [2]. OpenAI's own post notes it proved the "forced" variant, not the unforced version people picture when they hear "Millennium Prize," and independent mathematicians haven't verified it yet. Before you cite "AI solved a $1M problem" in a client deck, check whether the mathematician who almost got there first is calling it plagiarism.

OpenAI also shipped a Data agent inside ChatGPT Work — connect Redshift, Snowflake, BigQuery, or Databricks, ask a question in plain English, and it investigates what changed and builds a dashboard in Power BI, Tableau, or Sigma, no analyst or BI ticket required [3]. OpenAI says nearly all of its own product team already uses it daily. Business/Enterprise tiers only for now, but if "hire someone to build me a dashboard" has been a recurring line item, this is the vendor betting you'll pay to skip it.

Tooling shifts

n8n retired its old AI Workflow Builder — the one that generated a workflow once and stopped — for n8n Assistant, which builds the workflow on canvas, runs it, reads the actual node-by-node failure instead of declaring success, proposes a fix, applies it, and runs it again [4]. The output is still a normal n8n workflow: visible on canvas, editable by hand, logged on every run — not a black box. If you're still the one pasting error messages back into a chatbot to patch your own automations, this is the first version built to skip that step. Enabled by default on n8n Cloud; self-hosted needs Docker v2.36+. (If you're weighing n8n against the alternatives at all, that's the tool comparison I run through with clients.)

Instinct, the $2.5B-valued consumer AI assistant, now gives every user their own email address — layered on its existing 1Password login and Stripe integrations — so it can create accounts, contact businesses, and handle returns without touching your personal inbox [5]. It can sign up for a service, pay for it with your saved card, and get a return label without a human replying to a single email. That's the real test of whether you'd trust AI to act as you commercially, running at consumer scale before most SMBs have piloted the same authority internally.

Source: TechCrunch — Viral AI assistant Instinct now has its own email address

SMB angles

Zapier surveyed 518 execs at companies that already have a formal AI governance policy — not laggards, the ones supposedly doing it right — and 38% said unsupervised AI has already cost them time, money, or reputation: lost revenue, reputation damage, legal trouble, and 54% who spent real time redoing the AI's work [6]. Here's the gap that should worry you more than any single number: 49% of the same execs still say they personally trust AI to make most decisions without human review. That confidence-to-consequence gap, not the AI itself, is where your next incident comes from.

Infostealer malware is quietly draining Claude subscribers' token allowances by hijacking login sessions [7]. One consultant, Grant De Swardt, watched his usage jump from 45% to 55% with zero work done, got his $200/month account suspended, and received a partial refund before Anthropic reinstated him two weeks later — he never even got the warning email other affected users did. Anthropic will tell you your total usage. It won't give you an itemized breakdown to tell your own work from someone else's. If a usage spike on your Claude bill looks off, don't assume it's yours before checking your machine for malware — Anthropic isn't giving you the tool to tell the difference.

Adjacent to watch

Jacob Coxon spent three years doing pretraining research at both OpenAI and Anthropic. He quit this week saying the labs are "racing straight to self-improving superintelligence and gambling with our lives," citing Anthropic colleague Evan Hubinger's own estimate that AI kills everyone within a decade at "greater than 10%" [8]. Anthropic didn't respond to the resignation. It also didn't respond when The American Prospect reported that Anthropic's own security team runs a $180K–$230K "enterprise intelligence specialist" role to track activism, pays risk firm Samdesk for real-time protest monitoring, and had two managers describe on a podcast getting "60 minutes of advance notice" to reroute an executive around a protest [9]. Same company, same week: one researcher says the industry isn't watching the technology closely enough, while its own security team is closely watching the people protesting it. If your vendor-risk checklist for "AI companies I pay for enterprise seats" stops at data handling and never asks what their security team surveils, it's incomplete.

You don't need to trust the label "AI agent." You need to know whether anyone — including the vendor — can audit what it did after it acted alone.

Line up this week's stories by how much authority the agent actually had, and the shape gets obvious:

flowchart TD
  Start([This week's agent stories]) --> Draft["Reads + drafts:<br/>OpenAI Data agent<br/>builds the dashboard, you read it"]
  Start --> Act["Acts on its own:<br/>Instinct signs up + pays,<br/>n8n Assistant fixes its own runs"]
  Act --> Audit{Who can audit<br/>what it did?}
  Audit -->|No itemized log| TokenTheft["Claude token theft:<br/>victims can't prove misuse<br/>vs their own usage"]
  Audit -->|The vendor's own security<br/>team, watching activists<br/>not your usage| GSIS["Anthropic GSIS<br/>surveillance report"]
  Audit -->|Nobody yet,<br/>result unverified| NavierStokes["OpenAI's Navier-Stokes<br/>proof + credit dispute"]

Sources

[1] OpenAI — Navier-Stokes existence and smoothness problem — openai.com [2] TechCrunch — OpenAI fought dirty on career-making math problem, says NYU mathematician — techcrunch.com [3] OpenAI — Now everyone can put data to work — openai.com [4] n8n — Introducing n8n Assistant — blog.n8n.io [5] TechCrunch — Viral AI assistant Instinct now has its own email address — techcrunch.com [6] Zapier — AI and human approval — zapier.com [7] TechCrunch — Hackers are stealing Claude tokens from subscribers — techcrunch.com [8] TechCrunch — "Gambling with our lives": Anthropic researcher quits, warns against self-improving AI — techcrunch.com [9] The American Prospect — Anthropic's AI surveillance system to monitor activists — prospect.org

None of these individually change what you ship Monday morning. Together they say the industry handed agents real commercial authority — an email address, a debugging loop, a live dashboard — the same week a math result went unverified amid a credit dispute, a three-year insider quit calling the whole approach reckless, and a leading lab's own security team turned out to be watching protesters instead of its product's usage. Autonomy is the easy sell. Auditability is what nobody's shipping yet — so before you extend an agent's authority past "reads and drafts," find out who can actually check its work after it acts alone.

The short version

• OpenAI says 10,000 agents produced a Navier-Stokes proof in 88 hours for $22.5M — but it's the "forced" variant, unverified, and NYU's Tristan Buckmaster says OpenAI raced him to it using his own unpublished approach
• Instinct gave its AI assistant its own email address so it can sign up for and pay for services without you in the loop — a real-world test of commercial AI identity
• n8n Assistant is the first workflow builder that actually reads its own failures and fixes them, instead of declaring victory and walking away
• Zapier's own survey of 518 governance-minded execs: 38% got burned by unsupervised AI, yet 49% still say they trust it to decide without review
• Hackers are quietly draining Claude subscribers' token budgets via stolen sessions — and Anthropic won't give victims an itemized way to prove it wasn't them
• Same week an Anthropic researcher quit warning about reckless self-improving AI, reporting surfaced that Anthropic's own security team runs paid protest-surveillance on activists