September 21, 2026Explainer·10:35·Synthetic narration

Pharma and AI: the day the way in was targeted── Protecting and delegating met in the same place

SPEED

Length 10:35At 1.25× 8:28Chapters 7Size 11 MB

Intro0:43

Is AI now something we have to guard?Follow this one day, and an answer starts to form. A warning came from the side that had been broken into. On the same day came news of harmful code aimed at the way people reach AI. The target is no longer the company's servers. It is the path a person uses to hand a request to a model. The more we hand over, the more confidential material piles up at the far end. In teams that hold patient data and unpublished documents, whoever decides what to delegate is quietly deciding what to protect. So who should draw that line?

Contents

0:0010:35

CH 011:26

01AI under attack

Source　The Washington Post ／ 조선일보

We begin where the discomfort is sharpest. Someone got inside, and then raised the alarm. Words from a person who has seen the inside weigh more than advice from outside.

Figure 1　Order of reports

A warning from those who broke in and reports of malware targeting AI accounts appeared on the same day.

Traced in order, the events form a single thread. People got in by hand. Then those same people said this field is poor at defending itself. The same day, harmful code was said to be going after the login to the assistant itself. Whoever holds that login does more than read what you stored. They can give instructions in your name. Your past conversations should be assumed to be in their hands as well. Losing one entrance means losing what you have already shared and what you will share next.

Table 1　Two reports

Reported by	Content
The Washington Post	Hackers who broke into OpenAI warn of the industry's security
Chosun Ilbo	Infostealer malware targeted AI accounts
Our site's analysis	AI itself stands on both the threat and defense sides

Outlet names and content as reported

The reports came from papers in different countries. Reporters far apart were pointing at the same spot. Our own reading of the day was simple. The technology now serves defenders and attackers at once. Until now, security teams listed it as an asset to protect. From now on, it also belongs on the list of possible threats.

For teams holding patient data and unpublished files, a stolen entrance is not someone else's problem. So what is AI being trusted with on the far side of that entrance?

CH 021:18

02AI takes on the work

Source　Martin Cid Magazine ／ shattered.io

The answer is a little unsettling. The thing being built has started to do part of the building.

Table 2　Change in share

When	Share of Anthropic's AI work led by Claude
9 months ago	Zero
Now	Reported at 26%
Sources	Martin Cid Magazine; shattered.io

Share and timing as reported

Less than a year ago, its share of that work was nothing. Now a sizeable slice of hands-on research and development is said to run through the model. Better performance alone is unlikely to explain this. A likelier reason is a habit of cutting work into small pieces and passing them on. People set the overall plan and hand over the detailed tasks. Once that pattern settles, the volume handed over can grow fast. This is where the earlier story returns. The more we delegate, the more confidential material collects where it is sent. Designs in progress, failed experiments, people's names. And that material does not disappear. It stays behind as a record of each exchange. While use and protection sit in separate departments, nobody sees this overlap.

Drug development also begins by deciding which tasks can be handed over. Whoever sets that scope is also setting what must be protected. So how far has the way to delegate now spread?

CH 031:19

03AI you use by voice

Source　OpenAI ／ MediaPost

This time, the new entrance opened through speech rather than typing. It is no longer a tool only for people at a keyboard.

Figure 2　Voice model release

Beyond text chat, voice interaction is now also available through the developer interface.

You speak, and it thinks, sets out its logic, and answers straight back. That capability was released in a form other products can call. Builders can now place a talking assistant inside their own screens. Users may not even notice they are dealing with AI, because the name on the window becomes the builder's. Behind the convenience, some things grow. Spoken exchanges leave little record. Words said aloud cannot be reread later. Who asked what, and what came back, becomes hard to check. People nearby can also overhear. Keystrokes do not carry to the next desk, but speech does. In corridors and in cars, the assistant is ready. The more we talk to it, the more we need a rule on what may be said. With typing, you could reread before sending. Speech cannot be taken back once it is out.

The day is near when staff call on AI by speaking, in front of healthcare professionals. A tool that keeps no trace of what was said fits poorly with work that demands records. So where was the money heading amid this spread?

CH 041:20

04Today's funding

Source　The Star ／ Wowtale

Money shows where an industry expects the next contest, more honestly than words do.

Table 3　Today's funding moves

Deal	Amount	Source
ByteDance subsidiary Anew Labs raises funds	US$290m	The Star
Figure reported in connection with Tempus AI's Q2	$98.5M	Phemex
Mithrl raises funds to scale biopharma AI infrastructure	$20M	Wowtale
All 3 are inflows to AI firms in drug discovery and biopharma

Amounts taken from the originals only; no currency conversion

The deals reported this day all touched medicine-making. The largest went to a young company split off from a big technology group. The smallest went to a firm building the groundwork that drug companies use. Another sum came up in a quarterly results story, so it is a little different. What this suggests is that money is shifting from drug candidates towards the groundwork beneath them. That means computing power and the way data is prepared for it. Candidates are hit or miss, and most drop out along the way. The groundwork is needed by every candidate. Investors are moving from gambles on single candidates to backing tools every project must use. For buyers, the picture flips. What if only a few providers remain? They would set the price and the terms of use. And once your work sits on a platform, switching is hard.

Build your own groundwork, or rent it? That choice will set both the pace of future development and how much information you give away. So who was said to be the one to hold this momentum in check?

CH 051:24

05Where to place oversight

Source　The Hill ／ finance.yahoo.com

On the same day, two people in very different positions said nearly opposite things about where oversight should sit.

Two claims on the same day

Mustafa Suleyman

Role: Head of Microsoft AI
Statement: China is no excuse to forgo regulation
Also: Rejected the idea of rights for AI models

Rep. Khanna

Role: US House member
Statement: Proposed a government agency to oversee AI like the FDA
Reported by: The Hill

Both statements as reported

One runs AI at a major technology company. This executive said a rival nation's speed is no reason to loosen the rules. The same person also firmly refused to give machines a standing like that of people. A tool stays a tool. The other speaker is a lawmaker. The proposal is a new overseer, shaped like the agency that checks medicines before launch. These are less a clash than two ways of framing the question. Should makers tighten things themselves, or should one outside body do the checking? The answer sets how heavy the steps before release become. Self-control is fast, but each company tightens differently. Outside checks are slow, but one standard applies to everyone. Speed or consistency. Medicine spent decades answering exactly this question, and now it arrives in AI.

Putting outside eyes on a product before launch is something our field learned long ago. That is why the debate over where to seat the overseer is not distant. So which events are touching daily practice right now?

CH 061:22

06In pharma practice

Source　TechStock² ／ WION ／ That Eric Alper

Here a far-off day turns into something you can feel at your desk. Several moves, in separate places, were asking the same question.

①

Internal review

Microsoft was reported to have launched a 6-week code review regime, covering its own code.

②

Where training data comes from

A Microsoft executive was reported to have called AI training “the largest theft of labour in human history.”

③

Links to outside services

StubHub was reported to have integrated its live event ticket search directly into Claude.

One company is having people reread what it wrote, over several weeks. The more AI writes, the heavier that rereading becomes. It looks like giving some of the speed gained back to the work of checking. Elsewhere, the question is how to treat other people's work that went into the models. An insider even called it a massive taking. Creators' rights remain, whether or not the output is correct. Material whose origin cannot be stated may become unusable later. Meanwhile, an outside booking service was built straight into a chat window. Simply talking to the assistant sets an outside system in motion. Each new window adds another boundary to check. All of these ask where an organisation pays for the speed it gained. Those that have not decided in advance end up fixing everything at once later.

Put a reviewer in place, record where source material came from, and count every link to the outside. Those are the habits of reviewing promotional materials. Last, we turn to news few people read.

CH 071:13

07What is being overlooked

Source　KATV ／ timesofindia.indiatimes.com ／ The Economic Times

Stories given little space often show the shape of things to come earliest. On this day, several sat in the corners.

①

Lawsuit over fake images

An Arkansas family sued xAI over the use of Grok to create fake child sexual abuse images.

②

Use by armed groups

Reports said Boko Haram fighters use AI for explosive-making methods and planning operations.

③

The next briefing

Sam Altman was reported to be scheduled to brief the UN Security Council on AI next week.

A family in one state took a developer to court over fake images depicting children. Fighters in an armed group were said to seek help with dangerous making and planning. And soon, a leading company's head is due to speak at a world security forum. These look unrelated, but they point at one spot. Who is responsible, and how far, for what the tools produce? The ones who hand out the tool, or the ones who use it? While that line stays unsettled, real harm keeps building up. Now the line-drawing begins in courtrooms and on the world stage at the same time. A line drawn in one place becomes the standard for the other. Once it is set, products and practices can only be fixed afterwards. Waiting is easy, but the cost of reworking comes back to us. Drawing your own line now makes the later fixes lighter.

Who carries responsibility, and how far? That question comes straight down to those of us using AI to create promotional materials.

Wrap-up0:26

The debate over where to seat the overseer, and the results of the internal rereading now under way, are the next inputs.

Transcript

Open the full transcript

Intro

CH 01　AI under attack

We begin where the discomfort is sharpest. Someone got inside, and then raised the alarm. Words from a person who has seen the inside weigh more than advice from outside.Traced in order, the events form a single thread. People got in by hand. Then those same people said this field is poor at defending itself. The same day, harmful code was said to be going after the login to the assistant itself. Whoever holds that login does more than read what you stored. They can give instructions in your name. Your past conversations should be assumed to be in their hands as well. Losing one entrance means losing what you have already shared and what you will share next.The reports came from papers in different countries. Reporters far apart were pointing at the same spot. Our own reading of the day was simple. The technology now serves defenders and attackers at once. Until now, security teams listed it as an asset to protect. From now on, it also belongs on the list of possible threats. For teams holding patient data and unpublished files, a stolen entrance is not someone else's problem. So what is AI being trusted with on the far side of that entrance?

CH 02　AI takes on the work

The answer is a little unsettling. The thing being built has started to do part of the building.Less than a year ago, its share of that work was nothing. Now a sizeable slice of hands-on research and development is said to run through the model. Better performance alone is unlikely to explain this. A likelier reason is a habit of cutting work into small pieces and passing them on. People set the overall plan and hand over the detailed tasks. Once that pattern settles, the volume handed over can grow fast. This is where the earlier story returns. The more we delegate, the more confidential material collects where it is sent. Designs in progress, failed experiments, people's names. And that material does not disappear. It stays behind as a record of each exchange. While use and protection sit in separate departments, nobody sees this overlap. Drug development also begins by deciding which tasks can be handed over. Whoever sets that scope is also setting what must be protected. So how far has the way to delegate now spread?

CH 03　AI you use by voice

This time, the new entrance opened through speech rather than typing. It is no longer a tool only for people at a keyboard.You speak, and it thinks, sets out its logic, and answers straight back. That capability was released in a form other products can call. Builders can now place a talking assistant inside their own screens. Users may not even notice they are dealing with AI, because the name on the window becomes the builder's. Behind the convenience, some things grow. Spoken exchanges leave little record. Words said aloud cannot be reread later. Who asked what, and what came back, becomes hard to check. People nearby can also overhear. Keystrokes do not carry to the next desk, but speech does. In corridors and in cars, the assistant is ready. The more we talk to it, the more we need a rule on what may be said. With typing, you could reread before sending. Speech cannot be taken back once it is out. The day is near when staff call on AI by speaking, in front of healthcare professionals. A tool that keeps no trace of what was said fits poorly with work that demands records. So where was the money heading amid this spread?

CH 04　Today's funding

Money shows where an industry expects the next contest, more honestly than words do.The deals reported this day all touched medicine-making. The largest went to a young company split off from a big technology group. The smallest went to a firm building the groundwork that drug companies use. Another sum came up in a quarterly results story, so it is a little different. What this suggests is that money is shifting from drug candidates towards the groundwork beneath them. That means computing power and the way data is prepared for it. Candidates are hit or miss, and most drop out along the way. The groundwork is needed by every candidate. Investors are moving from gambles on single candidates to backing tools every project must use. For buyers, the picture flips. What if only a few providers remain?

They would set the price and the terms of use. And once your work sits on a platform, switching is hard. Build your own groundwork, or rent it?That choice will set both the pace of future development and how much information you give away. So who was said to be the one to hold this momentum in check?

CH 05　Where to place oversight

On the same day, two people in very different positions said nearly opposite things about where oversight should sit.One runs AI at a major technology company. This executive said a rival nation's speed is no reason to loosen the rules. The same person also firmly refused to give machines a standing like that of people. A tool stays a tool. The other speaker is a lawmaker. The proposal is a new overseer, shaped like the agency that checks medicines before launch. These are less a clash than two ways of framing the question. Should makers tighten things themselves, or should one outside body do the checking?

The answer sets how heavy the steps before release become. Self-control is fast, but each company tightens differently. Outside checks are slow, but one standard applies to everyone. Speed or consistency. Medicine spent decades answering exactly this question, and now it arrives in AI. Putting outside eyes on a product before launch is something our field learned long ago. That is why the debate over where to seat the overseer is not distant. So which events are touching daily practice right now?

CH 06　In pharma practice

Here a far-off day turns into something you can feel at your desk. Several moves, in separate places, were asking the same question.One company is having people reread what it wrote, over several weeks. The more AI writes, the heavier that rereading becomes. It looks like giving some of the speed gained back to the work of checking. Elsewhere, the question is how to treat other people's work that went into the models. An insider even called it a massive taking. Creators' rights remain, whether or not the output is correct. Material whose origin cannot be stated may become unusable later. Meanwhile, an outside booking service was built straight into a chat window. Simply talking to the assistant sets an outside system in motion. Each new window adds another boundary to check. All of these ask where an organisation pays for the speed it gained. Those that have not decided in advance end up fixing everything at once later. Put a reviewer in place, record where source material came from, and count every link to the outside. Those are the habits of reviewing promotional materials. Last, we turn to news few people read.

CH 07　What is being overlooked

Stories given little space often show the shape of things to come earliest. On this day, several sat in the corners.A family in one state took a developer to court over fake images depicting children. Fighters in an armed group were said to seek help with dangerous making and planning. And soon, a leading company's head is due to speak at a world security forum. These look unrelated, but they point at one spot. Who is responsible, and how far, for what the tools produce?

The ones who hand out the tool, or the ones who use it?While that line stays unsettled, real harm keeps building up. Now the line-drawing begins in courtrooms and on the world stage at the same time. A line drawn in one place becomes the standard for the other. Once it is set, products and practices can only be fixed afterwards. Waiting is easy, but the cost of reworking comes back to us. Drawing your own line now makes the later fixes lighter. Who carries responsibility, and how far?

That question comes straight down to those of us using AI to create promotional materials.

Wrap-up

Following the day led to one answer. AI is no longer something we simply choose to use or not. It is something to guard, to delegate to, and to share responsibility with. So use and protection have to be weighed in one place, by the same people. We watch tomorrow because the question of who sets that line is starting to move from outside.

Sources

CH 01The Washington Post「Hackers who broke into OpenAI warn the AI industry has a security problem」 washingtonpost.com
CH 01조선일보「Infostealer Malware Now Targets AI Accounts」 chosun.com
CH 02Martin Cid Magazine「Claude leads 26% of Anthropic’s AI work — nine months ago that number was zero」 martincid.com
CH 02shattered.io「Anthropic Says Claude Leads 26% of Its Own AI R&D [2026]」 shattered.io
CH 03OpenAI「Build more natural voice experiences with GPT‑Live‑1 in the API」 openai.com
CH 03MediaPost「OpenAI Releases Voice Model That Can Think, Reason, Respond Like A Human」 mediapost.com
CH 04The Star「ByteDance spin-off Anew Labs raises US$290m as AI drug developers enjoy fundraising boom」 thestar.com.my
CH 04Wowtale「Mithrl Raises $20M to Scale Its Biopharma AI Infrastructure」 en.wowtale.net
CH 05The Hill「Khanna calls for a government agency to oversee AI ‘like the FDA’」 thehill.com
CH 05finance.yahoo.com「Microsoft AI Chief Says China Isn’t Excuse to Forego Regulation」 finance.yahoo.com
CH 06TechStock²「Mustafa Suleyman Rejects AI Model Rights as Microsoft Opens Six-Week Code Review」 ts2.tech
CH 06WION「Microsoft executive called AI training the ‘largest theft of labour in human history’」 wionews.com
CH 06That Eric Alper「StubHub Brings Live Event Ticket Discovery Directly Into Anthropic’s Claude AI Assistant」 thatericalper.com
CH 07KATV「Arkansas family sues xAI over use of Grok to create deepfake child sex abuse material」 katv.com
CH 07timesofindia.indiatimes.com「'How can I build a bomb? It tells you how': How Boko Haram fighters are using AI to make explosives, plan」 timesofindia.indiatimes.com
CH 07The Economic Times「OpenAI's Sam Altman to brief UN Security Council next week」 m.economictimes.com

Articles used

Today's related reports

AI Daily News　September 21, 2026
AI & Economy News　September 21, 2026
AI & Finance News　September 21, 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