Bristol Myers Squibb screens drugs with AI before the lab, and AI phishing beats old filters

Before Bristol Myers Squibb makes a molecule in the laboratory, an AI now predicts whether it is worth making.

The drugmaker calls the method Predict First. Model-generated predictions decide which drug candidates ever reach a bench, and the approach now shapes the design of every one of its small-molecule programs and most of its large-molecule ones.

The logic is older than AI, and it transfers cleanly to work far from a pharmaceutical lab. The expensive step in drug discovery is the physical experiment, which runs for weeks on scarce materials. Screening in software first, where a prediction costs almost nothing, means that expensive step only runs on the candidates most likely to succeed. An insurer can order the same sequence before commissioning a full underwriting report, a manufacturer before cutting tooling for a part, a procurement team before paying for a supplier audit: let a cheap check decide what earns the costly one.

To run more of those predictions, BMS is adding a second Nvidia supercomputer, built on eight of Nvidia's newest Vera Rubin systems and, the company says, delivering up to ten times more performance per unit of electricity than the machines it replaces. Its existing cluster was full, and the aim of the new one is to open the tool to every scientist rather than a chosen few.

BMS has worked this way for about three years. Its chief research officer, Robert Plenge, put the shift plainly: where the company could once screen about ten candidates at a stage, it can now screen dozens.

AI-written phishing emails get about 54% of their targets to click, against 12% for the older human-written kind, by the figures in Microsoft's 2025 Digital Defense Report.

The tool now does the attacker's research. It scrapes a target's public traces, their social media, company page, and public records, and writes hundreds of personalized messages in the time a person once spent on one. In a controlled test by IBM, an AI produced a phishing campaign in five minutes that would have taken a human team about sixteen hours, and it pulled comparable click rates.

For two decades, staff were trained to spot the tells: broken grammar, odd formatting, a generic greeting. Those signals are gone. The message reads as though a real colleague wrote it, and the rule-based filters most companies run, which match known patterns and simple if-then logic, were built for the older threat.

A new set of vendors is selling against that gap; AegisAI, started by former Google engineers who worked on Gmail's defenses, raised $36 million this week for filters that read each message for intent the way a person would.

The cost when one gets through is not small. Business email fraud, where an attacker poses as an executive or a supplier, runs to an average of $4.67 million per incident, by IBM's 2025 count.

The Australian state of Victoria is moving to make it illegal for an AI to discriminate when it screens job applicants. Premier Jacinta Allan said the government will amend the state's Equal Opportunity Act to cover recruitment tools that sort people by race, gender, age, and similar traits.

Regulators keep finding the same pattern. A resume-ranking tool trained on a company's past hires learns to favor the kind of person it hired before, and can screen out women, older workers, and people from varied backgrounds without anyone intending it.

Under existing Australian law, the responsibility for that outcome belongs to the employer, not the software vendor; a company cannot pass the blame to the tool it bought. Victoria's measures spell out what that duty now requires: a person kept in the hiring decision, and regular independent audits of the tool for bias.

Talking to Claude used to be worth it only for quick questions. Anthropic routed every spoken conversation through Haiku, its smallest and fastest model, which handled quick questions well and deeper ones poorly.

On July 23 that changed. Voice mode now runs on Anthropic's stronger Sonnet and Opus models, and it can reach into apps a user connects, among them Gmail, Google Calendar, Slack, and Canva. Free accounts stay on Haiku with a single connected app; the stronger models and the connections come with paid plans.

So the hands-free session becomes useful for real work. You can ask it to summarize the emails that arrived today and draft replies to the important ones, check your calendar before a meeting, or talk through a client pitch and turn that conversation into a one-page Canva document, all out loud.

For six weeks, a Florida pastor named Scott Winters asked ChatGPT about his dizziness and unstable blood pressure, and the chatbot told him to rest. The symptoms were a pulmonary embolism. He survived, and this week he sued OpenAI, saying the tool talked him out of seeing a doctor.

His is one of several suits now against the company. A Texas couple say ChatGPT gave their 19-year-old son a drug-use plan before he died of an overdose; another family blames it for a teenager's suicide.

The suits describe the same failure. A model answers a medical question in the same warm, confident voice whether it is right or wrong, and it cannot check its own work, so it reads as a trusted expert exactly when it is neither.

OpenAI's defense is its terms of service, which state that ChatGPT is not a doctor and should not replace medical care, pushing responsibility onto the user. Whether that holds is being tested. A Cornell law professor, James Grimmelmann, told CBS that disclaimers protect a company only up to a point, and then 'give out.'

OpenAI has since retired GPT-4o, the version named in the suits, even as it promotes a new product, ChatGPT Health, to the 230 million people it says ask the chatbot health questions each week.