Today’s five stories ask who is in control as AI reads brain data, finds software flaws, attempts clinical work, places ads inside conversations and starts acting on what we say.

1

AI analyses millions of brain cells in 32 minutes. What has it actually discovered?

Thousands of AI agents have examined experimental data from millions of mouse brain cells, identifying possible connections to neurodegenerative disease. But generating promising hypotheses is not the same as establishing scientific discoveries.

Artificial intelligence has analysed a vast dataset of mouse brain cells in just 32 minutes, identifying possible biological connections that could inform future research into neurodegenerative diseases.

The work, announced on 21 October by OpenAI, PerturbAI and Scripps Research, used thousands of AI agents powered by GPT-Rosalind.

The agents examined an existing whole-brain CRISPR atlas, a dataset showing how disruptions to individual genes affect different types of cells across the mouse brain.

Rather than asking one AI model to interpret the entire dataset, researchers distributed the analysis among thousands of agents working in parallel.

Each investigated a particular genetic disruption, examined cellular responses and searched scientific literature for possible explanations.

The resulting reports were combined into a larger analysis connecting genes with observed cellular effects and potential biological mechanisms.

The researchers estimate that a comparable human-led analysis, examining the disruptions sequentially, could take months to a year.

Among the findings was a possible connection involving DCTN1, a gene associated with Perry syndrome, a rare neurodegenerative disease with Parkinson-like symptoms.

The AI agents linked the disruption of this gene to changes in cholesterol-related activity in brain cells. This generated a hypothesis about whether altered cholesterol distribution could contribute to neurodegeneration.

The relationship remains unconfirmed and requires further laboratory testing.

The researchers also compared the distributed approach with a single-agent method. In blinded assessments, human experts preferred the reports produced by multiple agents in 75% of comparisons.

The study illustrates how AI could help researchers examine more of the information generated by complex scientific experiments.

Yet its significance lies in identifying promising research questions, rather than proving new treatments or biological mechanisms.

Source: OpenAI

2

Are AI companies in charge of our digital safety?

OpenAI and Anthropic are uncovering weaknesses in software used across the internet, while OpenAI is proposing new rules for advanced cybersecurity tools. Are the companies developing the most powerful AI systems also becoming the guardians of our digital infrastructure?

The companies building some of the world’s most advanced artificial intelligence systems are taking on an increasingly influential role in cybersecurity. Their latest findings show how AI could strengthen digital defences, but also raise questions about the concentration of technological power.

On 21 October, OpenAI announced that researchers using its models had helped identify critical vulnerabilities in PHP, a programming language widely used to power websites.

The discovery involved the security company Calif, which used GPT-5.6 Sol, while cybersecurity company Wiz independently identified and reported the vulnerabilities. PHP’s maintainers led the fixes, with OpenAI and other organisations helping coordinate the response.

A day earlier, Anthropic reported that Claude had identified 493 issues across 25 open-source software projects.

The findings were not necessarily exploitable vulnerabilities. Anthropic’s research examined ways to reduce the amount of software code exposed to potential attackers, making some weaknesses harder to reach without first identifying every individual bug.

Both announcements demonstrate how AI could help secure software at a scale previously difficult to achieve.

But they also introduce a question about who possesses the resources to develop these capabilities.

Much of the software examined is publicly available. The advantage enjoyed by frontier AI companies therefore cannot simply be attributed to exclusive access to source code. It also reflects their ability to develop powerful models and deploy substantial computing resources.

This distinction matters as these companies begin to influence not only how vulnerabilities are found but also how defensive technologies are distributed.

On the same day as its PHP announcement, OpenAI published a policy blueprint for empowering cyber defenders. The company argues that there is a limited opportunity to strengthen defensive capabilities before comparable offensive AI tools become widely available.

Its proposals include stronger public oversight and wider access to advanced cybersecurity capabilities for legitimate defenders.

In September, OpenAI also announced $1 billion in subsidised access through its Daybreak programme, aimed at organisations with limited security resources.

The initiative could help expand access to advanced tools. However, it also illustrates how private technology companies are gaining influence over the distribution of capabilities that may become essential to public security.

Source: OpenAI

3

Medical AI can pass exams. Why does it still fail at clinical tasks?

Stanford researchers found that leading medical AI agents completed fewer than half the tasks in two demanding clinical and administrative benchmarks. The results expose the gap between knowing what should be done and actually doing it.

Artificial intelligence can achieve exceptional scores on medical examinations. But when asked to carry out complex tasks within healthcare systems, even leading AI agents frequently fail.

A Stanford University assessment published on 19 October found that the best-performing agents completed fewer than half the tasks in two benchmarks designed to evaluate clinical and administrative workflows.

In PhysicianBench, which measures clinical tasks carried out within an electronic health record environment, the strongest model completed 46% of tasks on a single attempt.

In HealthAdminBench, which evaluates administrative workflows, the highest-performing agent achieved 36.3%.

The benchmarks use different environments and evaluation methods, so the results are not directly comparable. Both nevertheless indicate considerable difficulties with multi-step tasks.

One finding was particularly revealing.

Among certain action-execution failures, 62% for GPT-5.5 and 73% for Claude Opus 4.6 involved agents that documented the correct action in a clinical note without actually carrying it out in the electronic health record.

An AI system could correctly state that a medical order should be placed, for example, but fail to enter the order.

That distinction could have serious consequences in a real clinical environment.

The Stanford researchers used synthetic patient data, allowing them to examine performance without putting patients at risk.

They also found that agents could complete individual stages of a workflow successfully while failing to finish the overall task.

In the administrative benchmark, one model completed 82.8% of individual subtasks but only 26.7% of whole tasks.

The findings suggest that improving healthcare AI will require more than expanding medical knowledge.

Systems must also demonstrate that they can act reliably across software environments and retain information throughout lengthy workflows.

Source: Stanford HAI

4

ChatGPT’s advertising system reads the conversation. What does that mean for users?

OpenAI has revealed how its advertising technology interprets conversations to identify relevant commercial messages. The company says ads remain separate from AI-generated answers, but the system offers a new model for commercial targeting.

OpenAI has published a technical explanation of how advertising works inside ChatGPT, revealing how a specialised AI model interprets conversational context to identify opportunities for sponsored placements.

In an engineering article published on 21 October, OpenAI researchers Ritendra Datta and Karthik Gopalratnam explain that advertising within a chatbot requires a different approach from traditional search advertising.

A search engine typically responds to a discrete query. In ChatGPT, a person’s intentions may become clearer over several exchanges, and sometimes across different conversations.

OpenAI says its advertising system uses a specially trained language model to interpret the available context and determine whether a relevant advertisement should appear.

The system can retrieve potential advertisements, evaluate their suitability and conduct an auction before selecting a sponsored placement.

According to OpenAI, the advertising infrastructure operates separately from the model generating ChatGPT’s answers.

The company says this separation prevents advertisements from influencing responses. It also states that no advertisement is shown unless the system determines that an appropriate placement is available.

This is an important architectural distinction, although the company’s published explanation is not an independent audit of its performance.

OpenAI reports that ChatGPT advertising reached an annualised revenue run rate of $1 billion fewer than 200 days after launch, with more than 100,000 advertisers having used the platform.

The development illustrates the commercial possibilities of advertising within conversational systems.

Unlike a conventional search query, a conversation can reveal changing intentions over time.

Someone discussing travel arrangements, for example, might reveal interest in accommodation only after several exchanges.

For advertisers, this offers opportunities for more contextual targeting.

For users, it raises questions about how conversational information is interpreted for commercial purposes and whether the distinction between assistance and advertising remains sufficiently clear.

Source: OpenAI

5

AI can now make calls and complete tasks. Who is responsible when it gets things wrong?

Mistral has introduced a new generation of voice AI designed to handle conversations and carry out actions through connected systems. The technology could transform customer services, but acting on spoken instructions brings new risks.

Artificial intelligence is moving beyond answering questions by voice. It is increasingly being designed to complete tasks while speaking to people in real time.

French AI company Mistral has introduced Voxtral 3, a family of speech models accompanied by software for building voice agents capable of interacting with connected services.

The system combines speech recognition with language-model reasoning, allowing developers to build assistants that can interpret instructions and take action during live conversations.

Mistral says its technology supports speech transcription across more than 50 languages, while its voice-generation models offer expressive speech in more than 35 languages.

The company describes applications ranging from customer service to booking appointments and handling administrative requests.

In one demonstration, an AI voice assistant retrieves information from a medical referral and offers to book an appointment after obtaining the caller’s confirmation.

The example illustrates a significant shift in what voice assistants are being designed to do.

Rather than simply providing an answer, an agent can interact with an external system and complete an action.

Mistral says Voxtral 3 can support phone calls through existing telephony infrastructure and operate through its European-hosted API or on an organisation’s own systems.

The option to host the technology independently may appeal to organisations concerned about retaining control over sensitive information.

But allowing voice assistants to carry out tasks introduces challenges that are different from those associated with conventional chatbots.

A transcription error could change the meaning of an instruction. An agent could misunderstand whether a caller intended to authorise an action or was merely asking about it.

A human customer service representative can also make mistakes. The question is how automated systems should be designed to recognise uncertainty and when they should seek confirmation.

Mistral’s announcement demonstrates the capabilities of the technology, but does not establish that every application built with it will reliably complete complex tasks.

Source: Mistral