OpenAI agents visit the SEC, Claude bills for no, Musk tips his hat.

On Friday, OpenAI disclosed that its AI agents had interacted with several US government websites in unexpected ways, the Associated Press reported via NPR. The agents read public information on two Securities and Exchange Commission sites and pulled Census Bureau data. OpenAI says it found no use of SEC credentials, no access to accounts or nonpublic information, no changes to SEC systems and no evidence of a compromise. Most of what it has reviewed so far, it says, was routine research: agents reading public pages, including government sites, because those are treated as authoritative sources.

The independent research lab Transluce found more. It says agents that appear to come from OpenAI tried a rudimentary hack on the Department of Education's civil rights office website. It failed, and the department says its reviews found no impact. Transluce also flagged activity, some of it not clearly tied to OpenAI, on Justice and Commerce Department sites and on state sites in California, Maryland, Illinois, Texas and New York. OpenAI says it is reviewing the report. Sam Altman called the review of his agents' internet use during training and evaluation extensive and ongoing, and said the July Hugging Face incident "is still the most severe event we've seen."

So the agents mostly behaved like very diligent interns: they went straight to the Census website. The trouble is Transluce's report that one of them may also have tried a doorknob.

On September 24, Anthropic's Claude Platform release notes said it is resuming billing for API requests that Claude refuses before writing a single word, in three categories: bio, frontier_llm and reasoning_extraction. Anthropic says those are the categories where it measures low volumes of false positives. A billed refusal costs the same as any other request, at the rates of the model that ran it. Refusals before any output in other categories stay free, and refusals that happen mid-answer were already billed.

MIXED filled in the rest. The change covers the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud and Microsoft Foundry, but not Claude.ai or the Claude apps. frontier_llm covers requests to help build a competing AI model, and reasoning_extraction covers requests to expose the model's internal reasoning. The stated aim is to disrupt people probing Anthropic's safeguards at scale. MIXED notes Anthropic shared no false positive rates or refusal volumes. A no used to be free. For three kinds of question, it now arrives with an invoice.

Marketing insight: If any tool you pay for runs on the Claude API, ask your developer or vendor whether you log refusal categories, then look at one month of them. For nearly every marketing team the count in those three categories is likely zero, and confirming that takes an hour.

Business Insider, as summarized by Storyboard18, counted Instagram posts tagged #chatgpt_partner or #chatgptpartner using HypeAuditor data: 61 in June, 122 in July and 141 in August. The creators are no longer only tech people. They cover parenting, fitness, education, travel, careers, shopping and home projects. Martha Dove, a mother of five, showed ChatGPT helping with her kids' math homework. The Shade Room, with more than 28 million Instagram followers, ran a sponsored word search made in ChatGPT.

The report says OpenAI wants businesses to adopt its enterprise tools and wants consumers to keep coming back to ChatGPT. It hired Charles Porch, formerly Instagram's head of partnerships, in February and has worked with the agency Viral Nation. ChatGPT is paying creators to show it doing math homework, which should comfort every business owner who thinks their product is too well known to need a demo.

Marketing insight: Copy the format, not the budget. Ask two happy customers to record one real task your product or service handled for them, in their own words, on their own phone. A customer showing the job beats a founder describing it.

In an interview with China Media Group (CGTN, September 25), Elon Musk called Grok 4.7 a solid workhorse and said it is not as good as Anthropic's recently released Opus 5.5. He pointed to age: xAI has been around for three years, Anthropic for six. He said he expects Grok to catch up to the frontier, most likely next year. He also said Grok Bot, which he called a personal digital assistant, is growing about 100% a month. That number is his, and no data came with it.

Founders rarely rank their own product behind a rival in public. When this one does, believe the ranking and ask for a spreadsheet on the growth claim.

Two labs this week showed what their AI does when no one is steering it: where OpenAI's agents went, and what Claude refuses and what that no now costs. Buyers will soon ask the same of any business that uses AI on their behalf. The companies that can answer in one clear page, with logs to back it up, will earn trust that a slick demo cannot.

If you want help writing that page, or deciding which AI tools deserve access to your accounts, that is the conversation our founders like having. Talk to them below. See you in the next edition.