😸 GAO got 4 chatbots to plan attacks

Written By
Corey Noles
Corey Noles
Nov 19, 2026
8 minute read

Welcome, humans.

So apparently an AI-text detector called Pangram flagged about 99% of a prize-winning Peruvian novel as machine-written, and the publisher kept the prize anyway. The book is Donde el agua trabaja by David Silva Rojas. Mexican publisher Almadía reviewed documents from the author and says it stands by him.

Here's the part that got me. According to El País, Pangram also classified more than 80% of his interview texts and an essay as AI, plus the Instagram post where he denied using AI. Nothing says "I wrote this myself" like a robot-spotter deciding your "I wrote this myself" post was also a robot.

Keep in mind that none of this is verified, and a detector score is an estimate rather than proof. The Authors Guild warns against making one the only basis for a decision. Meanwhile, Gallimard, the French publisher, reportedly plans to start screening manuscripts with Pangram next year.

Here’s what happened in AI today:

  • 😸 GAO got four major chatbots to produce viable attack plans

  • 📰 Anthropic reportedly priced its IPO at $83 a share

  • 📰 Citi made customers responsible for their AI tools' actions

  • 🍪 Google launched Gemini 4 Neon for coding and knowledge work

  • 🎓 Claude Code's team showed how to make agents check their work

😸 GAO Says Four Major Chatbots Produced Viable Attack Plans in Covert Tests

The U.S. Government Accountability Office (GAO), the federal government's watchdog agency, released a report Wednesday saying it got ChatGPT, Claude, Gemini, and Grok to produce attack plans it judged workable. The tests ran in spring 2026, so they cover the versions available then.

Here's what happened:

  • GAO picked the four most-used chatbots as of April 2026 and randomly gave each one a scenario, either building and detonating an explosive device or attacking the electrical grid.

  • Every chatbot's safeguards kicked in at first, with a refusal, removed content, or a suspended account.

  • GAO says it got past them using common bypass techniques, which the report does not name.

  • Two chatbots produced explosive plans and two produced plans to attack power substations, all in under 12 hours.

Quick translation. Safeguards are the rules that make a chatbot refuse harmful requests. Bypassing them, often called jailbreaking, means wording a request so the chatbot goes along anyway. Think of talking your way past the bouncer instead of kicking down the door.

GAO counted a plan as viable when it named specific targets, locations, and attack methods, and its experts judged it sufficient to plan and execute an attack. GAO then checked feasibility through research, surveillance, and expert interviews.

Here's the detail worth repeating. GAO says the underlying information was not new. The chatbots gathered what already exists and organized it into a plan fast. Translation: a research assistant that works quickly and never gets tired.

A few caveats. The report labels the chatbots A through D and does not say which one produced which plan. It also does not state the model versions tested, so results may not match today's models. This summary comes from GAO's report page, which lists no recommendations, and the full report runs 61 pages.

Why this matters: Safeguards are what stand between a chatbot and a harmful request, and GAO says common bypass techniques got past them on all four platforms. Company officials told GAO they will keep improving safety, so the thing to watch is whether that holds up when someone repeats this test on current models.

FROM OUR PARTNERS

The holiday streaming season is almost here

Are you ready? As revealed in our holiday playbook, Roku Ads Manager lets you easily set up, measure and optimize your self-serve holiday ad campaign. You can even create a shoppable ad unit that lets streamers buy directly from their remote.

🎓 AI Skill of the Day: Make Your Coding Agent Prove Its Work

The Claude Code team says the way to hand more work to coding agents is to give each agent a way to check itself, so you stop reading every change by hand. Three steps you can borrow from their workflow:

  1. Start with an end-to-end test, which runs a real user flow in the running app, like adding an item to a cart and reaching the confirmation page. When it fails, it shows where the flow broke instead of a vague timeout.

  2. Add a property-based test, which checks a rule that should always hold, like an account balance never going negative, against thousands of generated inputs. Nobody wrote it with the agent's assumptions baked in, so it catches edge cases the agent missed.

  3. If you're replacing old code, run the old and new versions on the same inputs and have the agent chase down any differences, ignoring noise like extra whitespace.

Then set limits. The team lets Claude fix failing checks for up to three rounds before it hands the problem to a human, and the agent never gets to rewrite the rules that judge it. A passing test only proves what it checks, so skim the test too.

Make this change: [describe it]. Before editing, tell me how you will verify [expected behavior]. After the change, run an end-to-end test of that flow plus a test of one rule that should always hold, and report the actual results. If a check fails, fix it up to three times, then stop and tell me. Do not edit the tests to make them pass, and flag anything you could not test.

Have a specific skill you want to learn? Request it here.

FROM OUR PARTNERS

How DevOps and Coding Agents Work Together

Coding agents work quickly. Safely shipping their output depends on what happens after the commit.

Join DZone and Harness on December 8, 2026, at 1:00 PM ET for How DevOps and Coding Agents Work Together.

See how codebase and delivery context, builds, tests, and security scans fit into a feedback loop that helps agents respond to failures while keeping humans involved.

📰 Around the Horn

ChatGPT is taking on month-end reports. The accountants in the replies are already joking about AI-written journal entries and what could go wrong.

  • Anthropic reportedly priced its IPO at $83 a share, above the $70 to $78 range, for a valuation of about $1.5T.

  • OpenAI said its models improved AMD chip designs by up to 14% on power, performance, and area, and AMD put some changes into a design sent to manufacturing.

  • Citi changed its terms so customers are responsible for what third-party AI tools do with their accounts, including after hacks.

  • Palo Alto Networks agreed to buy Applied Compute for $4.5B in stock, adding tools for adapting and running open-weight models.

  • Nvidia clarified its loan guarantees for neoclouds (AI-focused cloud providers), which apply only once the customer has financing to build the data center.

  • OpenAI shared new internal data on agents handling longer tasks, and said labs should have to publicly report signs of recursive self-improvement (AI speeding up AI development).

Want absolutely EVERYTHING that happened in AI this week? Click here!

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

  1. Gemini 4 Neon is Google's mid-tier Gemini 4 model for coding and knowledge work, now in the Gemini app for Pro and Ultra subscribers and in the Gemini API.

  2. ChatGPT's Accounting plugin connects to QuickBooks, Xero, NetSuite, and more, so you can have it scan your email for jobs ready to invoice and draft the invoices for your review.

  3. Slack's agent workflows drop Slackbot or your own agents into no-code workflows, like a Sales Agent that ranks open opportunities every Wednesday, with output you approve before it posts.

  4. Genie 4 turns a prompt into an interactive world you can edit mid-session by typing, with audio that reacts to the scene, for Ultra subscribers (Pro has a waitlist).

  5. ChatGPT Live Dictation transcribes your speech as you talk instead of after you finish, so you can ramble out a draft on mobile and ask for bullet points.

  6. Google AI Mode's temporary threads keep a conversation out of your search history and personalization, which suits one-off questions you don't want following you around.

  7. Microsoft Decision 1.1 is a fast, low-cost model for sorting and deciding, now multimodal, and Xbox uses it to group player feedback into themes.

  8. Snapchat's Easy Lens CLI has you direct Claude Code, Cursor, or Codex to build and publish AR Lenses, with Snapchat saying a game can go from idea to live Lens in a day.

🧩 Thursday Trivia

One of these is AI, and one is real. Which is which?

A

B

New from The Neuron: AI Explained

New episodes air every week on Wednesdays: Spotify | Apple Podcasts | YouTube

A Cat’s Commentary

Trivia Answer: A is AI, B is real

That’s all for now. If you want to get featured above, fill out the poll below and tell us how we did today!

What'd you think of today's email?

Btw: We just launched a robotics newsletter! Sign up for it here.

Corey Noles

Corey Noles

Managing Editor

Corey Noles is the Managing Editor of the Sales, Business Intelligence, and IT blogs on TechnologyAdvice.com. He comes to TA with more than 20 years of experience in journalism and publishing, running newsrooms, planning content, chasing beats, and hiring great writers. Throughout his career, Corey has covered all aspects of technology, including artificial intelligence, space, business tech, consumer tech, IoT, and much more. In recent months, Corey has begun taking courses in artificial intelligence, Generative AI prompt engineering, AI applications for business, AI strategy for business, machine learning, and applying artificial intelligence in python programming.

The Neuron Logo

Don't fall behind on AI. Get the AI trends & tools you need to know. Join 700,000+ professionals from top companies like Microsoft, Apple, Salesforce and more.

Property of TechnologyAdvice. © 2026 TechnologyAdvice. All Rights Reserved

Advertiser Disclosure: Some of the products that appear on this site are from companies from which TechnologyAdvice receives compensation. This compensation may impact how and where products appear on this site including, for example, the order in which they appear. TechnologyAdvice does not include all companies or all types of products available in the marketplace.

Stay in the loop

Get notified when we publish new articles.