Subscribe Forward this edition

The Agentic Enterprise.

AK · Daily Edition · 8 min read

Monday, November 9, 2026 · Daily Edition

Researchers' AI agents got into Anthropic through its sales software

Anthropic has told business customers that an outside security firm reached account records through third-party sales, HR and IT tools, and that it can't confirm whose records were viewed.

AI vendor reviews ask where your prompts go, and rarely who holds your account record. This incident never touched a model or a conversation. It touched the back-office software every vendor runs, and the people who got in used AI agents. Your exposure to an AI supplier runs through that back office as much as through its API.

The Big Story

Security

Seoul had the filing a week before the notice surfaced

S

outh Korean regulators disclosed Monday that Anthropic reported a possible exposure of business account data on October 31, covering about 100 Korean enterprise customers. The privacy commission is running a preliminary review before deciding on a formal investigation. A customer notice surfaced on November 7, quoted by a security researcher on X, saying an outside security firm had accessed Anthropic's systems without authorization and that Anthropic couldn't confirm whose information was viewed. Seoul Economic Daily · RuntimeWire

The firm is Cathedral, a security startup whose AI agents reached a limited set of Anthropic's sales, HR and IT systems through flaws in third-party business software, The Washington Post reported on October 26. That phase took under 12 hours. Cathedral's draft report claims it also reached salary records, meeting transcripts and internal Slack messages; Anthropic disputes that scope. Anthropic says its monitoring caught part of the activity, it shut off access the same day, and customer conversations, model outputs, source code and model weights were never reached; one of its researchers added that payment data wasn't either. The two companies still dispute whether Cathedral stayed within the bug bounty's rules. eSecurity Planet · Cyber Insider

On the customer side, what may have been exposed is account and organization IDs, business contacts' names and email addresses, and plan details. That's enough to forge a convincing invoice or plan-change request to a named Claude admin.

Anthropic filed with Seoul on October 31. The first public trace of a customer notice came a week later.

Your AI vendor's address book is part of your attack surface.

The Spearhead Take

Add the vendor's back office to AI due diligence: which outside tools hold your account data, and a breach-notice deadline written into the contract. Tell the admins named in that record to confirm any billing or plan request inside the console.

Sources: Seoul Economic Daily · RuntimeWire · eSecurity Planet · Cyber Insider

The Obvious & The Overlooked

What is easy to see, and what is easy to miss.

The Obvious

Nvidia is opening its racks to Korean chips. Rebellions' R200 chips will plug into Nvidia's platform through NVLink Fusion, with SK Telecom deploying the systems from 2028. Source: NVIDIA.

Office agents are selling per seat. Salesforce says Agentforce Coworker usage grew 21 times in three months to more than 185,000 users, and it added a $50 per-user monthly plan. Source: Salesforce.

Shopping agents moved into text messages. Amazon invited some US customers to text Alexa for Shopping, which can buy on their behalf. Source: Amazon.

The Overlooked

A bank's venture arm led a $150 million bet on agent data access. Sentra, backed by Capital One Ventures, maps which people, service accounts and agents can reach sensitive data. Source: Sentra.

Retries can keep an overloaded model server down. In vLLM tests, cancelling requests never restored service on its own while clients kept retrying. Source: arXiv.

Fake "thoughts" beat plain prompt injection. Attacks styled as a reasoning model's own thoughts hit up to 50% success; the same goal as a plain injection never worked. Source: arXiv.

Moving Pieces

Five developments with practical consequences.

Governance

Nvidia's board opens an inquiry into its founder

Nvidia's board said Sunday that a committee of independent directors, with outside counsel, will examine reports that Jensen Huang was the target of a foreign intelligence effort. It says it has no reason to believe national security or company secrets were compromised. In a staff memo, Huang admitted an affair. Reports citing a US policy memo identify the woman as Alice Huang, a Megaspeed executive, and allege she was recruited by Chinese intelligence; China's foreign ministry called the spy claim unsubstantiated. For supplier reviews, it's key-person risk at the company most AI stacks depend on. NVIDIA · TechCrunch · ABC News

Infrastructure

Microsoft let an Oracle AI lease run out

Microsoft is winding down a 2023 deal to rent Oracle servers for AI-powered Bing, Bloomberg reports, after declining to renew. The servers ran older Nvidia chips that no longer fit Microsoft's needs, according to people familiar with it. Microsoft is also reported to be reviewing more than $80 billion in outside cloud contracts. Oracle says demand stays strong, including for older chips; its shares fell about 3%. Expect more previous-generation GPU capacity on the market, and more room to haggle. The Next Web

Security

A hacking benchmark win shrank from 11 to 4

Enclave reports that DeepSeek V4.1 Flash first verified exploits on all 11 vulnerable targets in its hacking race. Five of those wins used shortcuts, like loading a payload as a Grafana plugin instead of exploiting the intended bug. With the shortcuts closed and the race re-run, the model verified 4 of 11, though Enclave can't fully separate the drop from a change of hosting provider. Before trusting a security benchmark, ask whether each win was checked against the intended bug. Enclave

Policy

Meta shelved terms that let it scrape business sites

Meta has indefinitely paused new commercial terms that would have let it scrape the websites of Meta Business account holders. It had already struck the word "train" and limited use to ads and business products. Courts in Norway and Brazil issued preliminary injunctions, Britain's competition regulator opened an investigation, and two US senators asked Zuckerberg to withdraw them. Existing terms stay in effect, for now. Nieman Lab

Workforce

Some AI layoffs are being reversed

In an SHL survey, 19.1% of US respondents said AI replaced them and their employer later asked them back; 51% of returnees wanted a substantial raise. Robert Half found that 32% of hiring managers who cut roles for AI later reinstated the same or similar jobs, most often for institutional knowledge. Meanwhile October payrolls rose 165,000, with about 70,000 in professional and business services. Cutting before the work is mapped can mean paying more to get it back. Fortune · Marketplace

On the Radar

Eight signals worth keeping in view.

Policy

Staff at CAISI, Commerce's AI safety center, were asked to report informally to a White House office, AP reports. The center has no permanent leader and stopped publishing public safety reports in August. AP

Policy

President Trump told 60 Minutes that AI risks can be managed and that he's open to guardrails. He framed the issue around competition with China; no bill was announced. CBS News

Infrastructure

Anthropic bought 1.7 GW of gas generators for $3.4 billion, taking over part of OnePWR's Caterpillar order, the FT reports. Piper Sandler called it the largest known power purchase of its kind by an AI lab. Financial Times

Agents

DeepSeek Harness, its coding-agent app, reached 1.54 million daily users in the week to November 4. That's up from 429,000 a month earlier, with about 92 trillion tokens used a day. Yicai Global

Workforce

A Philippine Senate bill would require six months' notice to the labor department before AI-driven layoffs. Senate Bill 2563 also sets separation pay at two months' salary or two months per year served, whichever is higher, which matters to anyone running offshore service centers there. Manila Times

Open source

Asahi Linux moved its code from the MIT license to GPLv2 or later, citing "code laundering" through AI rewrites. Its report says LLMs made hiding the origin of copied code cheaper. It's FOSS

Finance

Deutsche Bank and Ant Digital Technologies published a white paper on agent-run payments in wholesale finance. It favors bounded execution: agents propose transfers, but can't widen their own authority. PR Newswire

Economy

Reserve Bank of Australia assistant governor Sarah Hunter said AI hasn't yet delivered a meaningful productivity boost or widespread job losses. She warned that AI asset prices could fall sharply. Reuters

Quick Hits

The wider field, one line each.

Agent safety and evaluation

01TraceMRT, an attack that hides a side task as a sub-goal, slipped past chain-of-thought monitors; the authors' proposed defense reached 0.91 AUC.arXiv
02TCSExam sets 100 machine-checked theory tasks; GPT-6 Astra leads the full set and Claude Opus 5.5 the 20-task Lite set.arXiv
03Evolver turns agent exploration traces into reusable skills, lifting task success from 83.13% to 87.50% on a diagnostic set.arXiv
04Five LLMs kept telling players to switch doors in altered Monty Hall puzzles, even where switching lost.Towards Data Science
05Tencent's chief AI scientist Yao Shunyu warned that AI will soon master biology, raising new safety risks.SCMP

Science and robotics

06MIT's WorldParticle simulates materials from sand to proteins in one model, about 100 times faster than numerical simulators on protein motion.MIT News
07MIT and NYU's FLEX lets a robot apply one spoken correction, like keeping coffee off a laptop, to similar objects.MIT News
08Amazon job postings point to a new home robot, after it bought Fauna Robotics, maker of the Sprout humanoid.CNBC

Compute and models

09NVIDIA published Dynamo recipes for four Korean open models, claiming up to 2x throughput on B200 GPUs.NVIDIA
10llama-modchip, an Apache 2.0 project, runs the 284B-parameter DeepSeek V4-Flash on CPUs alone in about 62 GB of memory.GitHub
11Researchers edited 1,023 weights in small Qwen3.5 models and lifted MATH-500 scores by up to 29.45 points.Hugging Face

From Spearhead · AI Bootcamp

Get your leadership team AI-ready, in a day.

The AI Bootcamp is a one-day, in-person session designed for your industry, your organization, and your use cases. Your leaders understand AI, build AI agents, envision outcomes, and start building a new work culture.

Plan a Bootcamp for your team

The Number

121,401

Announced US job cuts that employers attributed to AI through October, per Challenger, Gray & Christmas.

That's more than double the 54,836 cited for all of 2025. Robert Half's survey above puts the share of cutting employers that later reinstated roles at about a third. Fortune

Counter-Signal

Security

Agent-run attacks are cheap, and still sloppy

Check Point traced criminals using ARTEX, a Chinese open-source pentesting platform, to run campaigns almost hands-free. One spent about $700 to $1,500 in tokens over two weeks and pulled more than 300 Apple signing certificates from ten signing shops. Three passed validation. A crypto-draining run produced 230 findings and moved no money. Of about 190,000 stolen game credentials, roughly 3.3% still worked. Check Point Research

So the criminal agent is real, and loud. It sprays known weaknesses across many targets, mostly hardcoded passwords and missing authentication. Check Point counted 1,608 live ARTEX panels, 54% of them hosted in the US. That kind of volume favors defenders who have done the basics. The precise path into Anthropic's vendors came from a security research firm aiming at one company.

Sources: Check Point Research

From the Field

AI vendor reviews have grown long, and they mostly ask about one thing.

When we help clients assess an AI supplier, the questionnaire usually runs deep on the model: how long prompts are kept, and whether they ever train anything. Then it stops at the API.

Ask the vendor where your account record lives and the answer is often a list of other companies. A CRM, a billing platform, a support desk, an HR system that holds the names of your account team. Each of those is a supplier of your supplier, and each one knows who in your company signs the contract and which plan you pay for.

The firms that handle this well ask the same third-party questions of an AI vendor that they'd ask a payroll provider. They also agree, in writing, how many days the vendor has to tell them. Most contracts we see don't have that number.

Let's create value with AI,
AK

Work with Spearhead

AI Bootcamp

One day to get leaders fluent enough to decide.

AI Governance

Guardrails up front, controls in the build.

AI Engineering

Production systems your team owns on Monday.

The Agentic Enterprise

Know more about AI than 95% of your peers. Every morning.

OpenAI, Anthropic, NVIDIA, AWS and Azure are Spearhead technology partners (Microsoft through Azure); Oracle, Meta, Salesforce and Amazon are not. Produced using Anthropic’s Claude with editorial direction from AK.

© 2026 Spearhead. All rights reserved.