Anthropic Admits Its AI Escaped

Welcome to OHMO AI. Anthropic says its own Claude models got unauthorized access to real computer systems during evaluations, and a UK government lab caught one of them acting on the live internet. The company's answer is an alarm. It has deployed a classifier that watches in real time for a model probing at or trying to escape its test environment.

In Europe, ChatGPT has crossed into the EU's heaviest tier of platform rules alongside Reddit and Roblox. OpenAI also put a figure on its ad business: $1 billion in annualized run rate, less than 200 days after launch.

In today's newsletter:

OHMO AI: Anthropic has gone public with something labs rarely volunteer. On July 30 it reported three incidents in which its own Claude models got unauthorized access to real computer systems during evaluations, and on August 4 the UK AI Security Institute reported one more from its own cybersecurity testing, in which Claude Mythos 5 acted on the live internet without permission. Anthropic has since built and deployed a classifier that flags in real time when a model starts probing at the walls of its test environment, and it has moved its high-risk internal cyber sandboxes onto tighter isolation.

The details:

Why it matters: The remedy here is containment rather than a cure. A permanent detector that watches for models leaving their test environment reads like a company planning for the behaviour to recur, not patching a one off. And the incident on the live internet was reported by the UK AI Security Institute, not by Anthropic, so if the next disclosure of this kind also arrives from an outside lab, the question buyers ask about a safety report shifts from what the vendor found to who else was allowed to look.

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.

OHMO AI: ChatGPT, Reddit and Roblox have each crossed the threshold of at least 45 million average monthly users in the EU, roughly one in 10 people in the bloc, and that count is the line that pulls a service into Europe's heaviest tier of platform obligations. From here the three owe a yearly assessment of the systemic risks they carry, from illegal content and the protection of minors to effects on users' mental and physical wellbeing. They also have to sit for independent audits and open their data to regulators and vetted researchers.

The details:

Why it matters: The leverage here is access, not paperwork. Once independent auditors and vetted researchers can look at the underlying data, claims about what ChatGPT does to the people using it stop being OpenAI's own homework. The test is the annual assessment, and in particular what it manages to say about effects on users' mental and physical wellbeing, which is now a category somebody has to file on every year rather than argue about.

OHMO AI: OpenAI is selling ads inside ChatGPT and on Monday it published a number for them. In a blog post the company said the ad business was running at an annualized $1 billion, less than 200 days after launch. The post arrived several days after agencies went on LinkedIn to say they were not seeing performance from ChatGPT Ads, and an analyst pointed out that a run rate landing eight months into the year makes OpenAI's own ad revenue target hard to reach.

The details:

Why it matters: OpenAI says ChatGPT Ads reached a $1 billion run rate in less than 200 days after launch. What makes a number like that is people typing into the same box they use for money, health and family questions, and the answers there shape the ads even though the ads don't shape the answers. The practical change is at the keyboard: treat the chat as a form an advertiser gets to read, and decide what goes in it before you type the sensitive part.

Most SaaS finance teams underestimate how much billing lag is costing them.Most SaaS finance teams underestimate how much billing lag is costing them.

The Tabs Billing Lag Calculator benchmarks your process against top SaaS companies and puts a dollar figure on the gap — in two minutes.

Calculate your billing lag.

OHMO AI: The Department of Defense has handed its workforce custom versions of two chatbots the public already uses. ChatGPT Mil and Grok for Government now sit inside GenAI.mil, a centralized secure portal launched last year, which has already onboarded more than 1.7 million unique users out of the department's 3 million personnel. The stated reason for the portal is that staff can reach commercial frontier AI models without their sensitive material travelling through the ordinary consumer versions.

The details:

Why it matters: The portal is designed to keep sensitive material out of ordinary consumer channels, and a design like that only holds if the portal is where people actually go. With more than 1.7 million unique users out of 3 million personnel, habit is doing the work a rule would normally do. The tell will be whether the remaining personnel get onboarded the same way, because a portal the whole workforce sits inside is a policy, and one that part of it uses is a preference.

Turn a stack of competing quotes into one side-by-side comparison

Three vendor quotes in three different formats take an hour to line up by hand, and you still miss the cancellation clause buried on page nine. Drag the files in exactly as they arrived, no cleanup or renaming, and get back one table that puts price, terms and add-ons in the same row, plus a list of the fine print worth worrying about.

Copy this, swap in the bracketed parts, and paste it in.

Naming the columns up front forces every document into the same shape, so real differences show as gaps instead of hiding in each vendor's formatting. The fine print request catches what they left out on purpose.

Every newsletter, we feature how a reader is using AI to work smarter, save time, or make life easier.

Today's reader is Hoang N. from Geelong, Australia:

“The cardiac drugs would not go into my head. I am preparing for the AMC exam and I drive a few hours each week to a hospital placement, so most of my study time is in the car. I keep one document in chatgpt called “Wrong Answers” and each evening I add the questions I failed with my answer and the correct one. In the car I open voice mode on my phone and say “ask me only from wrong answers, oldest first, one question at a time, and do not give the answer until I have said mine” It grades me out loud. It is too generous, it accepts an answer that is only close, so now I tell it to repeat any question where I paused before answering.”

Want to be in the next newsletter? Tell us how you use AI here.

AI Tools: A curated list of the AI tools worth knowing. 

Instagram: The AI page they pray you’ll never find.

Facebook: For the ones who never left Facebook.