November 7, 2026Explainer·9:39·Synthetic narration

The Day UK Regulators Made Firms Own AI Errors── Agents used in attacks, drug discovery AI deals, fewer hires

SPEED

Length 9:39At 1.25× 7:43Chapters 7Size 11 MB

Intro0:38

More and more, we hand a task to AI, and it plans the steps, picks up other tools, and comes back with the job done. That convenience leaves two open questions. When such a system makes a mistake, who takes the blame?

And as we hand over more, does the human role grow lighter?Today's reports offer clues from those who regulate, those who attack, those who invest, and those who hire. Drug companies handle patient data and product information. So one error weighs more here than in most industries. Where to draw the line between machine and human is the thread through today's news.

Contents

0:009:39

CH 011:21

01AI Errors Are the Company's Responsibility

Source　Computerworld

The first answer came from the regulators. When software that acts on its own fails, who cleans up afterward?

Table 1　UK Regulators' View

Item	What was reported
Who spoke	Four UK regulators
Applies to	Companies using autonomous AI agents
Laws applied	Existing competition, consumer protection and data protection law
Liability for errors	The company using the agent bears its errors

Computerworld report. Applies existing law, not a new law

Their answer was simple. The firm that chose to put the system to work also owns the result. No new rulebook is needed. The rules already in place simply apply. So the excuse that the machine did it alone fails from the start. As I read it, regulators shut the door on shifting blame.

This matters because such a system sets the order of its own work. Steps move ahead without a person checking each one. So mistakes tend to surface late. Everything that happens before anyone notices still counts against the user.

Picture this at a drug company. Suppose replies to inquiries, or drafts of materials, are left to the system. Suppose a wrong efficacy claim slips in. The company that sent it out carries the blame. So deciding who reviews the content, before handing work over, becomes a condition of use. Never release machine-written text that no person has read. That rule alone changes how much blame you carry.

Responsibility for machine errors now sits with the user. But how should a company prepare for outsiders who use these tools against it?

CH 021:13

02Attacks Using AI Agents

Source　Microsoft ／ Thomson Reuters

By the time defenders began using such tools, attackers were already using them. What were they asking the machines to do?

Figure 1　An Agent-Driven Attack

Confirmed by Microsoft Threat Intelligence. Attackers used an agentic attack framework and took further actions in the organization via the hijacked app.

The method itself is old. Find login details left where anyone can see them. Use them to get in, then keep acting inside. What is new is who does the tedious part, the searching. The machine took it over. It can likely scan far more places than people could, without rest. When attacking takes less effort, more companies may become targets.

The way in was a permission granted to an app. Nobody needs to steal a login. Whoever holds that permission can pass as a legitimate user. Activity after entry likely looks like normal use in the logs. That, I think, is why it is so hard to spot from outside.

Drug companies increasingly build software with outside developers. Are connection keys left sitting in shared code storage? Checking that is not only the IT department's job, in my view. For a company holding patient and physician data, a break-in leads straight to lost trust.

The entry point was a system outside the company. So would running these tools on your own machines make you safe?

CH 031:12

03Weak Spots in Locally Run AI

Source　Startup Fortune ／ Pasquale Pillitteri ／ ZDNET

Some firms now skip outside services and run their models in-house. But that brings a different risk.

Table 2　Moves in Locally Run AI

Move	What was reported
Low-cost model	DeepSeek V4.1 Flash, called a counter to the AI spending panic
Where spending goes	~90% of US firms' open-model spend to Chinese firms; Kimi 61%
Local substitute	Trying to replace Claude Code with a free local AI stack (ZDNET)
Flaw found	Malicious GGUF leaks local LLM memory. Ollama vulnerability

Headline-based reports. Article text not verified

Cheaper options have people questioning premium prices. Some stitch together no-cost tools to replace paid ones. Yet meanwhile, a weakness turned up in the software used to run models in-house. What you save on price returns as security work.

The money flow is a concern too. Most of what American firms pay for publicly released models is said to go to developers in one country. But that claim has not been checked against the original reporting. Using such models without asking where they come from worries me.

The weakness could be planted inside the very file that holds the model. Loading a file fetched from somewhere could leak data from the machine running it. In-house is no guarantee of safety.

Keeping research and patient data off outside servers is a sound reason to go in-house. But without checks on where files come from, the data you meant to protect may leak. If you start cheap, budget for checks from day one.

Here the story moves from defense to money. In what form is money flowing into machines that help make medicines?

CH 041:08

04Funding and Buyouts in AI Drug Discovery

Source　PR Newswire ／ TMX Newsfile

People who believe these tools can help make medicines show it with money. Today, the form that money took stood out.

Table 3　Capital Moves in AI Drug Discovery

Who moved	Form	Target and aim
Kos Biotechnology Partners	Investment	Company with a novel AI-driven drug discovery platform
iSpecimen	Acquisition	Foldlab AI products; expand into AI drug discovery, disease trends
Amounts	Not disclosed for either

What catches the eye is two different forms side by side. One is backing a young company in the hope it grows. The other is buying a finished tool outright. Neither deal came with a price tag. When the sums are hidden, who teamed up with whom becomes the main clue.

Why keep the sums quiet? The reports do not say. Perhaps they wanted to show the purpose before the price, but that is my guess. As readers, we can treat the missing figure as information in itself.

For drug makers, this means more potential partners. But a small firm's tools can change owners. After a sale, the terms of use or the handling of data may change. The more you rely on outside tools, the more you need to track who owns them. One option is to settle in the contract, in advance, what happens if the partner is bought.

Money is not only going to companies and tools. Research is also trying to change how the machines answer.

CH 051:11

05AI That Answers by Choosing Options

Source　arXiv

Machines usually answer in prose. Yet many business judgments mean picking one answer from a fixed set. A pick needs no long prose.

Figure 2　Asking a Decision Model

Each request pairs a 'state' with a 'question'. Three question types: pick one, yes/no, graded score. Preprint, not yet peer reviewed.

In the new approach, the system writes nothing. It returns a number for how well each candidate fits, and a person decides. The machine's job stops at supplying evidence for the decision. Because it never commits to a conclusion, people cannot skip thinking for themselves.

Generative vs Decision Models

Generative model

Returns: Reasoning and a written conclusion
Extracting the answer: Parsed from the output text
Cost and time: Cost and wait time add up

Decision model Drex Max

Returns: Only a probability per option
Extracting the answer: Not needed
Cost and time: Returns in far less time

Authors' report. Accuracy equal to or slightly below generative. Test conditions not given

Skipping the prose lets it respond much faster, the authors say. Accuracy is said to be about the same or a bit lower than prose-writing systems. What matters most to me is not speed but a consistent answer format. Numeric answers are easy to log and audit.

Here we return to the opening scene. Blame for a machine's mistake falls on the company using it. So keeping the conclusion with a person makes the line of responsibility easier to draw. Sorting adverse event reports, or rulings in promotional material review, look like a good fit. But this work has not yet been checked by other researchers. Any trial should start small.

So where in the daily work of drug makers has this technology already begun to enter?

CH 061:19

06Inside Pharma Operations

Source　Yahoo Finance Australia ／ミクスOnline ／ BioSpace

So far, we have mostly looked outside the company. Now we move inside the daily work of drug makers.

①

Sales platform

Veeva Systems was reported to be chosen as the core CRM for a major biopharma's AI rollout.

②

Joint trial in Japan

NTT, NTT DOCOMO BUSINESS and Igaku-Shoin began a joint trial with Ubie of a medical AI information platform.

③

Disease model research

Network Bio announced research with a leading medical center to build AI models of disease from biospecimens.

The technology is entering everyday core systems, not trial projects. It is being built into the system that holds sales records. A joint test has begun to build it into the channel that delivers medical knowledge. Sales and medical affairs both deal directly with doctors. On the research side, a hospital partnership aims to model how illnesses develop. These tools are moving into ordinary work.

Once inside core systems, errors spread differently. One rep's mistake affects one case. A flaw in the core system hits every rep nationwide at once. So decide early who sets it up and who checks it.

This is even more true when delivering medical knowledge. If doctors or pharmacists receive wrong content, patient safety is at stake. Under the regulators' view from earlier, the company that delivered it is responsible. The deeper these tools go, the heavier the reviewer's role. You also need a process for how far to correct an error once found.

The reviewer's role is getting heavier. So what is happening at the entry point where those reviewers are trained?

CH 071:09

07What Is Shrinking Is Hiring

Source　Resultsense ／ www.mhlnews.com ／ The Guardian

People often say machines will take our jobs. But the real change showed up somewhere unexpected.

①

Fewer hires

The Bank of England found AI is reducing hiring at UK firms but is not causing layoffs.

②

Staff cut by a tenth

47% of supply chain firms reportedly cut staff by 10% due to AI. Figures based on a headline.

③

People quitting

The Guardian reported white-collar workers quitting for fear of AI disruption. It says there is no evidence of mass layoffs.

What is falling is not the number of people let go, but the number brought in. The British central bank sees firms trimming new recruitment before cutting existing workers. Not recruiting is much harder to see from outside than firing. Meanwhile, some people are leaving their jobs out of fear. Workers' feelings are moving before the statistics do.

In some sectors, nearly half of firms have reportedly reduced their workforce already. But the figure comes from a headline, and the article has not been checked. A number spreading alone can become an excuse for cuts. The bigger the number, the more its source should be checked before use.

For drug makers, the problem is fewer places for young staff to gain experience. Reviewing materials and answering inquiries were how newcomers built judgment. Hand all of that to machines, and a decade from now no one may be able to check the content. To own a machine's mistake, you need people who can spot it.

Deciding what to hand to machines is also deciding how people will keep gaining experience.

Wrap-up0:23

Watch whether UK powers to investigate autonomous AI become a bill or a system.

Transcript

Open the full transcript

Intro

CH 01　AI Errors Are the Company's Responsibility

The first answer came from the regulators. When software that acts on its own fails, who cleans up afterward?Their answer was simple. The firm that chose to put the system to work also owns the result. No new rulebook is needed. The rules already in place simply apply. So the excuse that the machine did it alone fails from the start. As I read it, regulators shut the door on shifting blame.This matters because such a system sets the order of its own work. Steps move ahead without a person checking each one. So mistakes tend to surface late. Everything that happens before anyone notices still counts against the user.Picture this at a drug company. Suppose replies to inquiries, or drafts of materials, are left to the system. Suppose a wrong efficacy claim slips in. The company that sent it out carries the blame. So deciding who reviews the content, before handing work over, becomes a condition of use. Never release machine-written text that no person has read. That rule alone changes how much blame you carry. Responsibility for machine errors now sits with the user. But how should a company prepare for outsiders who use these tools against it?

CH 02　Attacks Using AI Agents

By the time defenders began using such tools, attackers were already using them. What were they asking the machines to do?The method itself is old. Find login details left where anyone can see them. Use them to get in, then keep acting inside. What is new is who does the tedious part, the searching. The machine took it over. It can likely scan far more places than people could, without rest. When attacking takes less effort, more companies may become targets.The way in was a permission granted to an app. Nobody needs to steal a login. Whoever holds that permission can pass as a legitimate user. Activity after entry likely looks like normal use in the logs. That, I think, is why it is so hard to spot from outside.Drug companies increasingly build software with outside developers. Are connection keys left sitting in shared code storage?

Checking that is not only the IT department's job, in my view. For a company holding patient and physician data, a break-in leads straight to lost trust. The entry point was a system outside the company. So would running these tools on your own machines make you safe?

CH 03　Weak Spots in Locally Run AI

Some firms now skip outside services and run their models in-house. But that brings a different risk.Cheaper options have people questioning premium prices. Some stitch together no-cost tools to replace paid ones. Yet meanwhile, a weakness turned up in the software used to run models in-house. What you save on price returns as security work.The money flow is a concern too. Most of what American firms pay for publicly released models is said to go to developers in one country. But that claim has not been checked against the original reporting. Using such models without asking where they come from worries me.The weakness could be planted inside the very file that holds the model. Loading a file fetched from somewhere could leak data from the machine running it. In-house is no guarantee of safety.Keeping research and patient data off outside servers is a sound reason to go in-house. But without checks on where files come from, the data you meant to protect may leak. If you start cheap, budget for checks from day one. Here the story moves from defense to money. In what form is money flowing into machines that help make medicines?

CH 04　Funding and Buyouts in AI Drug Discovery

People who believe these tools can help make medicines show it with money. Today, the form that money took stood out.What catches the eye is two different forms side by side. One is backing a young company in the hope it grows. The other is buying a finished tool outright. Neither deal came with a price tag. When the sums are hidden, who teamed up with whom becomes the main clue.Why keep the sums quiet?

The reports do not say. Perhaps they wanted to show the purpose before the price, but that is my guess. As readers, we can treat the missing figure as information in itself.For drug makers, this means more potential partners. But a small firm's tools can change owners. After a sale, the terms of use or the handling of data may change. The more you rely on outside tools, the more you need to track who owns them. One option is to settle in the contract, in advance, what happens if the partner is bought. Money is not only going to companies and tools. Research is also trying to change how the machines answer.

CH 05　AI That Answers by Choosing Options

Machines usually answer in prose. Yet many business judgments mean picking one answer from a fixed set. A pick needs no long prose.In the new approach, the system writes nothing. It returns a number for how well each candidate fits, and a person decides. The machine's job stops at supplying evidence for the decision. Because it never commits to a conclusion, people cannot skip thinking for themselves.Skipping the prose lets it respond much faster, the authors say. Accuracy is said to be about the same or a bit lower than prose-writing systems. What matters most to me is not speed but a consistent answer format. Numeric answers are easy to log and audit.Here we return to the opening scene. Blame for a machine's mistake falls on the company using it. So keeping the conclusion with a person makes the line of responsibility easier to draw. Sorting adverse event reports, or rulings in promotional material review, look like a good fit. But this work has not yet been checked by other researchers. Any trial should start small. So where in the daily work of drug makers has this technology already begun to enter?

CH 06　Inside Pharma Operations

So far, we have mostly looked outside the company. Now we move inside the daily work of drug makers.The technology is entering everyday core systems, not trial projects. It is being built into the system that holds sales records. A joint test has begun to build it into the channel that delivers medical knowledge. Sales and medical affairs both deal directly with doctors. On the research side, a hospital partnership aims to model how illnesses develop. These tools are moving into ordinary work.Once inside core systems, errors spread differently. One rep's mistake affects one case. A flaw in the core system hits every rep nationwide at once. So decide early who sets it up and who checks it.This is even more true when delivering medical knowledge. If doctors or pharmacists receive wrong content, patient safety is at stake. Under the regulators' view from earlier, the company that delivered it is responsible. The deeper these tools go, the heavier the reviewer's role. You also need a process for how far to correct an error once found. The reviewer's role is getting heavier. So what is happening at the entry point where those reviewers are trained?

CH 07　What Is Shrinking Is Hiring

People often say machines will take our jobs. But the real change showed up somewhere unexpected.What is falling is not the number of people let go, but the number brought in. The British central bank sees firms trimming new recruitment before cutting existing workers. Not recruiting is much harder to see from outside than firing. Meanwhile, some people are leaving their jobs out of fear. Workers' feelings are moving before the statistics do.In some sectors, nearly half of firms have reportedly reduced their workforce already. But the figure comes from a headline, and the article has not been checked. A number spreading alone can become an excuse for cuts. The bigger the number, the more its source should be checked before use.For drug makers, the problem is fewer places for young staff to gain experience. Reviewing materials and answering inquiries were how newcomers built judgment. Hand all of that to machines, and a decade from now no one may be able to check the content. To own a machine's mistake, you need people who can spot it. Deciding what to hand to machines is also deciding how people will keep gaining experience.

Wrap-up

Put today's reports together, and the answer is clear. However independently the machines act, people and companies own the results. So the more work we hand over, the more we need reviewers, and places to train them. In pharma, those places are material review and inquiry handling. Will this thinking take firmer shape in national rules?

That is why we keep following the news tomorrow.

Sources

CH 01Computerworld「Enterprises must take blame for their AI agents’ mistakes, say regulators」 computerworld.com
CH 02Microsoft「Threat actors use agentic AI tooling to find and abuse exposed secrets」 microsoft.com
CH 02Thomson Reuters「The autonomy problem: Why AI is outrunning network architecture」 thomsonreuters.com
CH 03Startup Fortune「DeepSeek's cheap V4.1 Flash model undercuts the AI spending panic」 startupfortune.com
CH 03Pasquale Pillitteri「Bleeding Llama: Malicious GGUF Drains Your Local LLM Memory」 pasqualepillitteri.it
CH 03ZDNET「I tried replacing Claude Code with a local AI stack that’s completely free – how it went」 zdnet.com
CH 04PR Newswire「Kos Biotechnology Partners Announces Investment in Fathom Therapeutics to Advance Novel, AI-Driven Drug Discovery Platform」 prnewswire.com
CH 04TMX Newsfile「iSpecimen Expands Into AI-Powered Drug Discovery and Disease Trend Monitoring With Acquisition of Foldlab AI Products」 newsfilecorp.com
CH 05arXiv「Towards Foundation Decision Models: Drex」 arxiv.org
CH 06Yahoo Finance Australia「Veeva Systems (VEEV) Becomes Core CRM Choice For Major Biopharma AI Rollout」 au.finance.yahoo.com
CH 06ミクスOnline「NTT・NTTドコモビジネス・医学書院医療AI情報プラットフォームの実装へ Ubieと共同実証を開始」 mixonline.jp
CH 06BioSpace「Network Bio Announces Research Collaboration with a Leading Academic Medical Center to Build AI Models of Human Disease from Biospecimens」 biospace.com
CH 07Resultsense「Bank of England: AI is cutting hiring at UK firms, not causing layoffs」 resultsense.com
CH 07www.mhlnews.com「47% of Supply Chains Reduced Headcount by 10% Due to AI」 mhlnews.com
CH 07The Guardian「Fearing AI disruption, white-collar workers are quitting their jobs」 theguardian.com

Articles used

Today's related reports

AI Daily News　November 7, 2026
AI & Economy News　November 7, 2026
AI & Finance News　November 7, 2026

About

The narration is synthetic speech. The content draws only on this site's articles from the same day. Each edition passes seven plain-language checks before publication. That means known obstacles to comprehension are held below threshold — it is not a guarantee of comprehension.

← All editions