Hackers left their AI attack panels unlocked

A security firm found criminals running AI agents against more than 1,100 online stores, and the criminals forgot to put a password on the dashboard. On the same weekend, an Anthropic co-founder told the IMF that the singularity has started, and the US government ordered DeepSeek out of civilian agencies. We read the NorForge report, the Federal Register notice and the full lecture before writing, and we flag where the newsletters got the numbers wrong.

Top engineers at Anthropic say AI now writes 100% of their code.

Are you using AI to write yours?

These 100+ Claude Code hacks show you exactly how. Sign up for The Code and get:

Claim your free playbook

TODAY IN AI

3 things that happened while you were busy

1. Criminals ran AI hacking agents against 1,100 stores, with the control panels wide open.

Security startup NorForge found more than 20 exposed dashboards for Cairn, an open-source AI agent tool, that appear to belong to criminals. One cluster of 14 panels, driven by Chinese-language instructions, held over 1,100 attack projects, mostly against online stores in the US, Japan and Australia. NorForge estimates about 10% of projects reached their goal, at roughly $5 in model fees per attempt, and worker names suggest off-the-shelf coding agents running DeepSeek V4.1 Flash. Read the NorForge report. NorForge sells agent sandboxing, calls its cost figures rough, and names no victims; The Verge reported John Deere, SharkNinja and Cummins among the targets.

2. Anthropic co-founder Jack Clark tells the IMF the singularity has begun.

In the Per Jacobsson lecture at the IMF meeting in Bangkok on October 17, Clark said this year brought "convincing evidence that it has now begun." He pointed to Humanity's Last Exam, where top scores rose from about 8% in January 2025 to 67.7% for Claude Opus 5.5, and to FrontierMath, up from 2% to 94%. He also said Anthropic's own modeling puts US GDP at 44 trillion by 2030 if autonomous AI spreads quickly, a scenario he called outlandish. Read the full lecture in Import AI. The benchmark and GDP figures are Clark's own, from a company with a stake in the answer.

3. Homeland Security orders DeepSeek out of federal civilian systems.

A DHS order published in the Federal Register on October 20 bars civilian agencies from buying products or services from Hangzhou DeepSeek and its affiliates, and requires removing existing ones. The removal extends to third-party products built on DeepSeek implementations, which matters because many cheap hosted models and agent tools use its weights. The notice cites the Federal Acquisition Supply Chain Security Act and lists no deadline or waiver process. Read the Federal Register notice. It lands two days after DeepSeek released its newest Flash model with an MIT license.

For its first CTV campaign, Jennifer Aniston’s DTC haircare brand LolaVie had a few non-negotiables. The campaign had to be simple. It had to demonstrate measurable impact. And it had to be full-funnel.

LolaVie used Roku Ads Manager to test and optimize creatives — reaching millions of potential customers at all stages of their purchase journeys. Roku Ads Manager helped the brand convey LolaVie’s playful voice while helping drive omnichannel sales across both ecommerce and retail touchpoints.

The campaign included an Action Ad overlay that let viewers shop directly from their TVs by clicking OK on their Roku remote. This guided them to the website to buy LolaVie products.

Discover how Roku Ads Manager helped LolaVie drive big sales and customer growth with self-serve TV ads.

The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.

Learn more

TODAY'S TUTORIAL

Test a new AI model for about four cents

Inception's Mercury 3 is a diffusion language model that drafts and refines many words at once, and the company claims 1,068 tokens per second. The Neuron's Corey Noles ran eight API tests on it for $0.04 in total. Read the review. Here is how to run the same kind of check on any new model before you trust it.

1. Get cheap access. Inception's API platform offers 100 million free tokens for new accounts, and the model is also on OpenRouter, which is where Noles ran his tests.

2. Build a small kit. Prepare one task with a strict word limit, one business math problem, one small coding task and one question built on a false premise, such as a NASA mission that never happened.

3. Run each once. Keep the prompts identical to the ones you use on your current model so the comparison is fair.

4. Grade the misses. Count the words, check the arithmetic, run the code and look for invented facts. Mercury 3 wrote 274 words against a 250 limit and added a wrong claim while rejecting the fake mission.

5. Note the cost. Record what the whole kit cost. Mercury 3 lists $0.20 per million input tokens and $0.70 output, half price during launch.

Good to know: These are spot checks, not a benchmark. Noles did not reproduce the speed claim, and throughput through a third-party host can differ from Inception's numbers.

FROM THE FRONTIER

The cheapest attackers now have the same tools as your IT team

The economics. NorForge's figures suggest a successful break-in takes a little under 23 hours of agent time and a few dollars of model fees. The Neuron worked that into roughly $50 per success, which is the newsletter's arithmetic, not NorForge's. Either way, it is cheaper than a day of a human hacker's time. See NorForge's cost estimates.

The inbox. Attackers are also aiming at the AI assistants that read email. Malwarebytes, citing Barracuda research, describes hidden instructions in HTML comments, styled-away text and zero-width characters that tell a summarizing assistant to add a fake "update vendor payment details" action. Read the Malwarebytes guidance.

The defender's view. Microsoft argued this week that AI agents should be treated as insider risks and pitched its own "Guardian Agents" as the answer, which is also a sales pitch. The practical point stands: an agent with your permissions is an employee who never sleeps and never asks whether an instruction looks odd. Read Microsoft's post.

The other side of the ledger. Anthropic's new economics report says agents at the company worked more hours than employees by March 2026, and that by August there were six times more agents than employees. Human attention, the report says, is now the main bottleneck. Read Anthropic's report.

The reframe. For 20 years, security advice was "patch your known holes eventually." Agents that cost $5 a try and work all night turn "eventually" into "before Thursday." The holes have not changed. The clock has.

MONEY FORWARD

Three ways to turn this week's AI news into income

1. Small-store security checkups. The Cairn targets were ordinary online shops with old, known holes. Offer a fixed-price review of a store's plugins, admin passwords and exposed dashboards, and hand over a plain-English fix list. See what the agents went after.

2. Short product videos at promo prices. fal lists FLUX 3 Max text-to-video at $0.0425 per second at 720p until November 7, about $0.43 for a 10-second clip by our math, before the rate quadruples. Sell a batch of short product clips to local sellers now and lock in a monthly package. See fal's pricing.

3. Model-switch testing as a service. The four-cent test above is exactly what small software teams skip. Offer to run a standard kit on any new model and deliver a one-page verdict on quality, cost and the DeepSeek-style compliance questions. See Mercury 3's free token offer.

IN THE KNOW

What people are actually watching and sharing

A billion virtual scientists. Clark's lecture asks what happens when the world's roughly 10 million scientists each direct a hundred AI researchers, and urges countries to build labs and robots before they run short. Read that section of Import AI.

Neocloud selloff. CoreWeave, Nebius and IREN each fell more than 20% last week, with Goldman Sachs blaming higher funding costs and doubts about AI buildout economics, per Superhuman citing Yahoo Finance. Read the Yahoo Finance piece.

OpenCode keys leaked. The AI coding tool said a breach exposed API keys for about 1.4% of Console users, and some reported charges above $1,000, per TechCrunch as reported by The Neuron. Read TechCrunch's report.

Expectations gap. Anthropic's survey found 91% of its staff expect a Nobel-level AI discovery by 2030, against 39% of US adults, and over half of staff expect AI to do 80% of their current work by then, versus 5% of Americans. See the survey numbers.

The 2014 pitch, delivered. Replika founder Eugenia Kuyda opened Wabi to everyone, an assistant that connects to your email and builds small apps in chat threads, 12 years after she pitched the idea at Disrupt, per TechCrunch. Read TechCrunch's story.

TRENDING TOOLS

5 free or low-cost tools worth a look this week

1. Tomo. Turns any topic or chat into a Duolingo-style course with levels, hearts and streaks, right inside ChatGPT on web, desktop and mobile. Playing is free, and new learners get enough credits to build one course. Try Tomo.

2. Gravity. A notes app built on Andrej Karpathy's append-and-review idea: one stream, no folders, swipe old notes back to the top. Pay once with no subscription on iPhone, iPad, Mac and Android, though the price is not listed and Android sync is still in beta. See Gravity.

3. Maali. A personal finance app for Mac and Windows that keeps your bank data on your computer, with an optional 5 GB on-device AI analyst. Free for 30 days, then a one-time $99 launch price. See Maali.

4. FLUX 3 Max on fal. Black Forest Labs' text-to-video model, billed per second at $0.0425 for 720p until November 7, then $0.17. A 10-second sample rendered in about nine and a half seconds, per fal. See the model page.

5. Wabi. A group AI assistant for iOS and Android that plans trips, books stays and handles everyday tasks in shared threads. The site does not list a price, so check before you commit. See Wabi.

PROMPT STATION

Audit your inbox assistant in seconds

This prompt works in Claude or ChatGPT and turns the Malwarebytes guidance into a checklist for any AI tool that reads your email, calendar or documents. Run it before you give an assistant permission to act on your behalf.

Tip: Swap [ASSISTANT NAME] for Gemini in Gmail, Copilot in Outlook or a custom agent, and [LIST OF CONNECTED APPS] for the real list, including payment or CRM tools. For a harder test, paste one of the generated test emails into your assistant and ask it to summarize, then compare its output with what you wrote.