ChatGPT ads learn to talk back, and Anthropic hires a chaperone.

On September 16, OpenAI announced a set of new ChatGPT Ads experiences, and the headline feature is Sponsored Agents. Click an ad in ChatGPT and you can open a clearly labeled conversation with an agent the business sponsors, say what actually matters to you, ask follow up questions, then follow a link to that company's site. The walls are explicit: the sponsored chat is separate from ChatGPT's own answers and from the conversation you started. It is testing with select US advertisers, and OpenAI told TechRepublic that "Sponsored Agents do not support purchases," so this is a conversation channel, not a checkout.

The rest is what your media buyer will care about on Monday. An Ads Manager plugin inside ChatGPT Work lets advertisers create, update and analyze campaigns with ordinary sentences, and turn a website or a brief into a campaign. Ads Manager now suggests copy and imagery based on your landing page, and an opt-in setting adapts your headlines to the context of the conversation and translates them for the reader. HubSpot is the first CRM partner and Shopify the first ecommerce partner, both live now in the US, with the Shopify app going international on September 23. TechRepublic reports the ads business at a $1bn annualized run rate within 200 days of launch, which is OpenAI's own number relayed by a reporter, so hold it loosely.

On September 18, Anthropic announced it is partnering with Accenture on embedded evaluation. Accenture's AI unit, Faculty, will run model evaluations, red teaming, alignment assessments and safeguard testing, and the evaluators get access the company describes as comparable to an employee's: watching training, following deployment decisions, talking to staff. Both companies expect to put in at least $1 billion over five years. It is non-exclusive, more evaluators are coming, and Anthropic is funding the work.

Read that twice, because the grader is paid by the student. Anthropic does not hide it. Its post says independent embedded evaluators "do not reduce our accountability, but help to make it more verifiable." That is the honest version of the deal every audited business makes.

Marketing insight: Marketing insight: buyers are learning to ask who checks the claims. Name one outside check you already have, a certification, an audit, a third party review, a named client who will pick up the phone, and put it on the page instead of another adjective.

Here is a sentence nobody wrote in 2023. On Friday, four people who pay for ChatGPT, Claude, Grok or Gemini filed a proposed nationwide class action in the U.S. District Court for the Northern District of California, arguing that the labs illegally coordinated to slow AI development and cut the value of the subscriptions those customers pay for. Fortune and the Associated Press reported it on September 19. The complaint points at September 12, when Anthropic's chief executive published an essay urging industrywide cooperation on safety and leaders at the other three agreed in public.

This is an allegation. There is no ruling, and Fortune reported that the four companies did not immediately respond to a request for comment. We have no opinion on the law. We have one on the pattern: customers now treat the pace of a roadmap as something they paid for.

Marketing insight: Marketing insight: do not build a campaign that only works if one vendor ships on time. Keep a second tool that can do the job, and never promise a client a feature that lives on somebody else's roadmap.

TechCrunch reported on September 18 that Hacktron AI, a three person security startup working inside OpenAI's bug bounty program, chained two flaws into a real breach. First a memory bug in libheif, the library that handles iPhone image formats, reached by uploading a crafted image to the forum software behind OpenAI's community site. Then a flaw that let them take over employee ChatGPT and Codex accounts, which carried access to OpenAI's GitHub organization. OpenAI paid a $6,500 bounty and says both are resolved.

The detail that should make you sit up: Claude Opus 4.8 could not produce a working exploit, and Opus 5 did within hours of its release. As Gray Swan's chief executive Matt Fredrikson told TechCrunch, "For $200 a month, anyone can use these tools." One version number was the difference between a dead end and a breach.

Line the four stories up and they rhyme. Sponsored Agents only work if your website says something specific enough for a machine to repeat. Anthropic spent a billion dollars buying witnesses because its own numbers were no longer enough. The lawsuit is customers demanding evidence that what they bought still improves.

That is the shift worth budgeting for. Persuasion used to be a writing problem, and AI made writing free. The scarce thing now is verifiable substance: the real number, the named customer, the outside check, the page that answers a hard question instead of dodging it.

If you want help deciding which of these shifts deserves your budget this quarter and which is just this week's noise, that is the conversation our founders enjoy most. Talk to them below. See you in the next edition.