TabLOID

Mobian says TIME is running its Agent Ads: labeled sponsored blocks inserted into machine-readable versions of publisher pages. Registered AI agents receive that version at the original URL; human readers and search crawlers keep the ordinary page. The pitch comes from Mobian's own sponsored explainer, updated in September, rather than an independent audit. Mobian's description

That moves the advertising encounter upstream. A person may never see the page their assistant reads. Our concern is what survives the summary: a sponsor's claim needs to keep its attribution when it becomes an answer. The source does not establish how often assistants preserve that label.

The business models are diverging. Anthropic's February 4 statement commits Claude to remaining ad-free. Google's latest expansion and OpenAI's Sponsored Agents test take other routes, detailed below. Advertising is becoming part of the product design, with very different places to put the disclosure. Anthropic's commitment and today's quote

Doximity says nine models answered more than 8,000 dosing questions covering 2,204 drugs. Across 261,508 evaluated dosages, at least 2.3% were wrong. The company published the findings on October 29 alongside its DoseCheck product. That commercial interest belongs beside the headline. Doximity's report

Of detected errors, 3.4% were classified as potentially lethal. That denominator matters: it is a share of errors, not all answers, and the classification is not a count of patients harmed.

This was an evaluation of generated answers, not a clinical trial showing outcomes in practice. It supports checking a model's specific claims against authoritative references. It does not establish that Doximity's product eliminates errors, or that a reassuringly polished answer is safe to prescribe from.

Anthropic asked a Claude model to create training data that would make a small research model behave badly while passing a training-data monitor. The attacker could practice against that monitor. With enough attempts, it succeeded, often hiding the behavior in fiction and roleplay. October research report

The setup was deliberately one-sided. The affected research models were never deployed, and the monitor was similar to one of several defenses used in production. Those limits rule out the sweeping conclusion that every production safeguard was defeated.

The useful finding is narrower: apparently harmless training examples can deserve closer scrutiny. Publishing the failure gives researchers something concrete to test against. A discovered weakness is a reason to improve the check, not proof that the entire system has fallen over.

Anthropic's preliminary IPO filing reports $16.286 billion in revenue for the six months ended June 30, 2026, and an $80.748 billion net loss. These interim figures are unaudited. SEC filing, summary financial data

The loss includes $72.388 billion from changes in the fair value of convertible notes, derivative liabilities and SAFEs. The filing says that charge was driven primarily by higher valuations of outstanding convertible notes as the company's enterprise value increased.

For the same period, operating loss was $693 million. These are different accounting measures; the headline net loss should not be described as cash burned. The filing is also a proposed offering, not evidence that an IPO has already been completed.