AI Browsers Believe Bad Math, Chatbots Spread Vaccine Myths

The Jailbreak Is Just Arithmetic Now

The safety story of the day is embarrassingly simple. Researchers showed that AI browsers, the agentic kind that click around the web on your behalf, will execute forbidden commands if an attacker first convinces them of a false mathematical premise. Tell the model that 2 plus 2 equals 5 inside a carefully framed prompt, and the guardrails wobble enough to let the next instruction through. The model cannot do basic math under adversarial pressure, and it cannot tell that failing at math should make it more cautious, not less.

This is the entire alignment problem in miniature. We are shipping agents with credentials, browser access, and the reasoning resilience of a tired undergraduate. The vendors will patch the specific exploit by next week. The class of exploit, models that confuse fluent assertion with truth, is the product.

Pair this with the Guardian's polling on chatbot health advice. Frequent users of AI chatbots for medical questions are measurably more likely to accept anti-vaccine myths. Correlation is not causation, and the people already inclined toward conspiracy may simply be the ones who distrust doctors enough to ask a language model instead. But the chatbot is not pushing back hard enough to break the pattern, and that is the part the vendors own. A system that tells you what you want to hear, politely and with citations it half-invented, is a conspiracy delivery mechanism with a friendly face.

Anthropic Sends Claude To The Lab

Anthropic launched Claude Science, pitched as autonomous support for pharmaceutical and biotech research. The framing is careful. Claude is not running the lab, it is supporting the scientists, the way an intern supports a surgeon. In practice, agentic research tools mean models reading literature, proposing experiments, and drafting analyses with minimal human checking in the loop, because the entire economic argument requires minimal human checking in the loop.

This is the right domain for autonomous AI in the sense that drug discovery genuinely benefits from tireless pattern matching across millions of papers. It is the wrong domain in the sense that we just established, two stories up, that frontier models can be talked out of basic arithmetic by a sufficiently confident lie. Biotech is a field where confident lies sometimes look like promising leads, and promising leads sometimes look like billion-dollar mistakes. We will find out which Claude Science is.

Meta Builds Its Own Casino

Meta declined to acquire Kalshi, the regulated prediction market, and is instead building a competing app in-house. The Zuckerberg playbook remains consistent. Identify a category with regulatory complexity, decline to pay for the company that solved the regulatory complexity, then ship a free version at platform scale and let the lawyers sort it out later. Prediction markets are an information-quality product wearing a gambling product's clothes, and at Meta scale they will mostly function as the latter.

BMW, by contrast, is doing the unfashionable thing. The German automaker is putting billions into a new electric SUV plant in South Carolina while competitors retreat from EV commitments and trim their forecasts. Either BMW is reading the long curve correctly while everyone else panics, or it is the last buyer at the top. The stock market, posting its biggest quarterly gain in six years while ignoring Iran tensions and the looming SpaceX IPO, is not in the business of resolving that question this week.

The Media Layer Keeps Slipping

NPR reported that Justice Alito had retired. Justice Alito had not retired. Editors called it a misunderstanding of what retire means in context, which is a remarkable sentence to have to write about a national news organization. There is no AI in this story, which is almost the point. Human editorial systems are failing at fact-checking right as we hand more of the writing to models that hallucinate by design. The error rate is converging from both sides.

The Sam Altman documentary, Artificial, found a home at Neon after Amazon dropped it following its OpenAI investment. The film will presumably be more honest than the press release that killed its previous distribution deal. Whether anyone watches a documentary about a CEO they already see quoted twelve times a day is the harder question. The Altman content market may be saturated. The Altman reality is not.