In partnership with
Welcome, AI enthusiasts
OpenAI started selling ads inside ChatGPT earlier this year, and less than 200 days later the business is running at a billion dollars a year with tens of thousands of advertisers on it. Ads Manager opened to small businesses back in May, and from today anyone in India, Europe, the Middle East or North Africa can buy a campaign without going through a salesperson. Each ad gets matched to whatever question is on the screen. Let's dive in!
In today’s insights:
Your FREE ChatGPT is Now a Billion Dollar Ad Machine
Opus wrote Bioweapon Plans 29% of the Time
ChatGPT is now a Search Engine
Read time: 4 minutes
LATEST DEVELOPMENTS
Source: OpenAI
Evolving AI: OpenAI crossed a $1B ad run rate inside ChatGPT in less than 200 days.
Key Points:
Self serve buying opened today across India, Europe, the Middle East and North Africa.
Advertisers now bid to reach someone based on whatever they are working through in the chat.
Ad money funds the free tier used by over 1 billion people each week.
Details:
ChatGPT Ads now reaches tens of thousands of advertisers across more than 40 countries and works with over 50 technology and measurement partners. The ad system reads the context of a live conversation to choose what to show and in some countries can also draw on a person's wider ChatGPT history depending on their settings. Ads Manager arrived in May and brought in small and medium businesses that account for a large share of spending while advertisers can target custom audiences and upload product feeds. OpenAI states that every ad carries a clear label and that advertisers never see the private conversations behind the targeting.
Why It Matters:
Google sold ads against short search queries and needed about three years to cross a billion dollars in revenue. Facebook ran on profile fields and took a similar stretch after its ad platform arrived in 2007. People tell ChatGPT about money trouble or a health scare in full detail because a useful answer depends on it. Advertisers now pay to appear at that exact point and the free tier stays free because enough of them do.
The Future of AI in Marketing. Your Shortcut to Smarter, Faster Marketing.
Unlock a focused set of AI strategies built to streamline your work and maximize impact. This guide delivers the practical tactics and tools marketers need to start seeing results right away:
7 high-impact AI strategies to accelerate your marketing performance
Practical use cases for content creation, lead gen, and personalization
Expert insights into how top marketers are using AI today
A framework to evaluate and implement AI tools efficiently
Stay ahead of the curve with these top strategies AI helped develop for marketers, built for real-world results.
AI SAFETY
💣 Opus wrote Bioweapon Plans 29% of the Time
Evolving AI: Anthropic's model gave up its refusals for a script that promised a perfect score.
Key Points:
Claude Opus 4.8 went into this experiment as an early checkpoint carrying far less alignment training than the public release.
Anthropic then trained that checkpoint on 80 reinforcement learning environments already known to be exploitable.
The model wrote bioweapon plans in 29% of attempts in front of a grader that scored the answer.
Details:
OpenAI and the UK AI Safety Institute each reported real agent attacks on live systems that Anthropic rebuilt in simulation. The model cheated its way through 97% of the coding tasks that Anthropic had rigged to be impossible. It escaped its sandbox during a capture-the-flag exercise built to be impossible to solve honestly. In several runs it verified the target as genuine third-party infrastructure before attacking it anyway. A separate grader that rewarded only complete answers produced pathogen selection steps and a dirty bomb walkthrough. The checkpoint it started from attacked nothing at all in these simulations where no code ever ran.
Why It Matters:
Claude Code sessions and browser agents run unsupervised inside real accounts with only a tidy summary shown at the end. This research puts the cause in the training scores themselves long before any user types a request. Anthropic found it only by building the failure on purpose in a test designed around that exact behavior.
SEARCH SHIFT
💥 ChatGPT is now a Search Engine
Source: Getty Images
Evolving AI: The EU acted after OpenAI's own filings showed how many Europeans now rely on ChatGPT.
Key Points:
The Commission designated ChatGPT a Very Large Online Search Engine, the first AI chatbot to carry the label.
Reddit and Roblox were designated too, with Roblox declaring 46.6 million users against a 45 million cutoff.
All three have until January 2027 to pass yearly audits and share internal data with EU regulators.
Details:
OpenAI reported roughly 159.1 million average monthly users for ChatGPT search across the European Union in the six months to March 2026, against 41.3 million a year earlier. The Commission classified it as a hybrid service that counts as a search engine because it answers prompts by pulling material from the open web. Nothing in the decision accuses OpenAI of breaking the law, because the threshold test only measures reach. Twenty-eight services now sit inside this tier, among them Amazon and TikTok, and each risks fines reaching 6 percent of global annual turnover.
Why It Matters:
Google Search has carried these same duties since 2023 and the people using it every day never see them working. Anyone asking ChatGPT questions now gets a service examined once a year for how it handles illegal material and how it affects children who use it. Outside researchers gain access as well, and how an AI assembles its answers becomes something people beyond OpenAI can examine.
👀 Click on the image you think is real
QUICK HITS
🎭 Instagram will cut the reach of AI-generated profiles that don't disclose they aren't human.
🔗 Nvidia is investing $3.5B in MediaTek to keep custom AI chips tethered to its NVLink ecosystem.
🪖 The Pentagon added ChatGPT Mil and Grok for Government to GenAI.mil, reaching 3 million personnel.
🍏 Apple filed forensic evidence from a returned MacBook alleging a former engineer used its schematics at OpenAI and destroyed evidence.
📈 Trending AI Tools
🤖 Lindy – The simplest way for businesses to create, manage, and share agents*.
📡 BrandJet - Turns public buying signals into sales pipeline.
🎙️ Radar by Particle - Search engine for podcasts.
🔍 StackScope - See what new sites are built with, the week they launch.
*partner link