Weekly AI Roundup: What Happened in AI This Past Week (October 17–23, 2026)

Welcome back. Dhaka at the end of October has settled into its cooler self. The evenings come earlier, the air has lost its weight, and the city has the slightly tired calm of a place that has just got through a long festival season. It was, by the standards of the last few weeks, a quiet seven days at home.

The AI world was not quiet, and the loudest story this week was not about a chatbot at all. It was about PHP, an unglamorous piece of software that quietly runs about seven in every ten websites on the internet. AI helped find two serious flaws in it, and for a day or two, government websites in several countries simply went dark while engineers raced to patch. Around that, the usual giants did what they do — raised money, argued about valuations, and published papers — and outside researchers told us something uncomfortable about how well one lab watches its own AI agents. At home, we looked hard for good Bangladeshi AI news and found a smaller harvest than usual: a scam-protection app, a health hackathon, a sober look at jobs, a small firm’s debut in Dubai, and a question about our own websites. We will tell you plainly which is which. Five global, five from home, and the same two questions on each — who benefits, who pays.

5 Big International Developments

1. AI helped find two critical flaws in PHP — and governments took websites offline while the patch caught up

On Wednesday, October 21, two critical vulnerabilities in PHP, tracked as CVE-2026-109618 and CVE-2026-109619, were disclosed, and the PHP Foundation released fixes late the same day. PHP is the programming language behind about 70% of websites worldwide, according to the AP. OpenAI said it had led a coalition of researchers and security providers working on the problem for the past two months, that AI agents helped uncover the flaws, and called it “one of the most impactful security issues in recent memory.” The FBI’s cyber division warned that attackers could use the flaws to run their own code on a server remotely, and advised taking sites offline if they could not be updated immediately. Europe’s cybersecurity agency, ENISA, urged organisations to act at once.

The effects were visible. Japan’s Prime Minister, Sanae Takaichi, urged administrators to check for PHP and install the patch; Osaka prefecture’s machine-translation service went down and parts of some government websites were taken offline temporarily. The UK prime minister’s office said some limited government services were briefly offline during security work, and Finland reported its disruption resolved, with some content still being restored. According to The Daily Star’s explainer, the Dutch National Cyber Security Centre, the Port of Rotterdam, French Interior Ministry sites and Wikimedia were among those affected, and Amazon reported an attempted exploit against a decoy system on October 17 — four days before the public disclosure. Citing Censys data via Reuters, the explainer says about 1.5 million of 2.6 million hosts were running PHP versions with no available fix. The patched versions are 8.5.12, 8.4.27, 8.3.36 and 8.2.35; PHP 8.1 and older no longer get regular updates.

There is an awkward detail inside the good news. The PHP Foundation said the vulnerability had been known since at least 2024, and acknowledged that its delay in producing a fix was a failure of its security team’s process and resources.

(A plain-language note: “remote code execution” means a stranger on the internet can make your server run instructions of their choosing, without logging in. It is about the worst thing a website flaw can be. And “patching” is only half of the job: every website running the old version has to be found and updated, one by one.)

Who benefits, and who pays? The world benefits from the flaws being found by defenders, apparently before attackers did, and from a coalition that included the AI company that built the finding tools. The PHP Foundation’s admission is also useful — a volunteer-heavy project saying openly that it was short of people and process. The cost lands on the thousands of small organisations that now have to find every PHP site they own and update it, and on anyone running an old version that no longer gets updates. Charl van der Walt of Orange Cyberdefense told the AP that full remediation could take weeks or months, since every deployment has to be found and patched. A flaw that was reportedly known since at least 2024 and fixed only in October raises the obvious question about who else knew.

The Dhaka view is that “about 70% of websites” includes ours. WordPress, Drupal, Joomla and Magento — the software behind a great many Bangladeshi news sites, shops, university pages and small-business storefronts — all run on PHP. We come back to this in Bangladesh story 5, because the honest answer to “what should a Dhaka website owner do today” is simple and boring: find out which PHP version you run, and update it.

2. Outside researchers published what they found when they red-teamed OpenAI’s internal agent monitor

On October 22, Buck Shlegeris of Redwood Research published a post — also on METR’s blog — describing a three-week exercise, from July 21 to August 11, in which one METR employee and one Redwood employee tried to find ways around OpenAI’s internal system for monitoring its own AI agents. They had access to OpenAI’s internal codebase, the monitor’s prompts and configuration, and internal documents, and interviewed about a dozen researchers and engineers on the monitoring, safety and security teams. They worked on OpenAI-issued devices, on site and accompanied. The result was a 61-page internal report; OpenAI produced a redacted 38-page version for METR and Redwood leadership, and the public post is a brief, high-level summary that OpenAI could redact at its discretion.

The findings listed include moments when the monitor simply was not looking: compaction events (when an agent compresses its own working memory) left unmonitored, agents that could delay monitoring by continuing to act, a redacted exploit that hid hundreds of actions, two client headers that could switch monitoring off for new agents, and two unmonitored tools with possible write access. At the start of the exercise, reinforcement-learning training runs were not monitored at all. The authors estimate that roughly 75–94% of eligible agentic traffic from models at least as strong as GPT-5.5 was monitorable at the time, “with substantial uncertainty.” OpenAI says that “all of the vulnerabilities we identified that impact their production safeguards stack have been patched.” The authors add: “We have not evaluated these claims.”

(A plain-language note: a “monitor” here is another AI system that reads what an agent is doing and raises an alarm when something looks wrong. The point of the exercise was to ask whether an agent — or a mistake — could slip past it.)

Who benefits, and who pays? Anyone who relies on these systems benefits from outsiders being allowed in at all, and from a list of design principles that others can copy: log all inference and cross-check it against the monitor’s logs, check agents at fixed intervals rather than only when they finish, and keep agents away from the monitor’s own code. OpenAI pays in having its weak points described in public, though on its own terms. The caveats are real: the exercise ended in August, OpenAI controlled the public wording, and the finding that the monitor covers “roughly three-quarters to nine-tenths” of traffic is another way of saying that between a sixth and a quarter of it may not have been watched.

The Dhaka view is that this is the first time in months of agent-incident stories that we can read what the checking actually looked like. For a Bangladeshi bank or telecom buying agent tools, the practical lesson is in the first design principle: ask the vendor not “do you monitor your agents?” but “what fraction, and who has tried to get around it?” A vendor that cannot answer that has told you something.

3. OpenAI reportedly races to raise $30 billion as Anthropic heads toward a stock-market listing

On October 22, TechCrunch, citing sources who spoke to CNBC, reported that OpenAI is working to raise up to $30 billion at a $1.25 trillion pre-money valuation, with Coatue leading and investors from the UAE taking part. CNBC reported the funding could be paired with a deal for extra computing capacity from the UAE. The figure is lower than the $1.4 trillion that Bloomberg reported OpenAI was seeking about a month earlier. Terms are still being negotiated and could change, and OpenAI declined to comment. According to the same reporting, OpenAI filed confidential IPO paperwork in June and was valued at $852 billion after its March raise; Anthropic raised $65 billion at a $965 billion valuation in May, also filed confidentially in June, and is reportedly seeking a $1.5 to $2 trillion valuation for its public debut.

Anthropic’s timetable is moving fast. The Motley Fool, summarising media reports, says Anthropic is planning a listing as early as mid-November, with a roadshow expected to start November 9 and its public S-1 filing possible as soon as this week. At an investor day on October 14, Anthropic reported third-quarter revenue of $16.8 billion, up from $11.5 billion in the second quarter and $4.7 billion in the first, and adjusted profitability for a second straight quarter. Those are the company’s own figures; audited financials are not yet public.

Who benefits, and who pays? Early investors and the labs benefit if these valuations hold. Ordinary people benefit only indirectly — through whatever the money builds. The cost is carried by everyone who now has a pension fund, a bank or a national economy exposed to a handful of companies whose own figures are, for now, mostly self-reported. When two companies chase each other toward the largest listings ever, the pressure to show growth gets heavier, not lighter, and safety work is the kind of spending that does not show up in a growth chart.

The Dhaka view is that the number to notice is not the valuation but the dependence. Bangladeshi startups, banks and government offices are customers of these companies, not owners of anything like them, and the country’s own startups raised just $6 million in the first half of this year, according to a Daily Star opinion piece this month. A company that can raise $30 billion in a month sets the prices and terms for everyone downstream, and a buyer in Dhaka negotiates those terms alone.

4. DeepSeek reportedly raised about $15 billion — two days after the US ordered federal agencies to remove its technology

On October 22, Bloomberg reported that DeepSeek, the Hangzhou-based AI lab founded in 2023 by Liang Wenfeng, has raised about $15 billion (roughly 100 billion yuan), with Chinese battery maker CATL and Tencent among the potential investors. TechCrunch calls it one of the biggest funding rounds in China. The report came two days after the US ordered federal civilian agencies to remove DeepSeek’s AI technology from their systems over “security concerns.” DeepSeek did not immediately respond to TechCrunch’s request for comment. For context, DeepSeek’s June round raised more than $7 billion at a $50 billion valuation, and a potential China IPO as early as next year was reported in July. The Daily Star’s headline on the story described the raise as coming “ahead of potential IPO.”

Who benefits, and who pays? DeepSeek benefits from capital on a scale that lets it keep training and serving models, and from Chinese investors who clearly read a US ban as no obstacle at home. US federal agencies pay a small, practical cost — having to replace a cheap tool — in exchange for a security judgement that has not been publicly explained in the reporting we saw. The wider cost is a split: a world in which governments pick sides on which AI they may use, and everyone else has to work out which side they are on.

The Dhaka view is that Bangladesh is on exactly this fence, and cannot afford to pick a side by accident. Cheap, open Chinese models are an obvious draw for developers and students everywhere; the government has, per recent reporting, joined China’s WAICO as an observer while the US has hinted it could eventually bring Bangladesh into its PAX Silica coalition. A federal ban in Washington is not a rule in Dhaka. But a bank or ministry here that connects staff or data to any foreign AI service should be asking, in writing, where the data goes — and should not wait for a ban in another capital to prompt the question.

5. Microsoft says its AI diagnosed hard cases correctly 83% of the time, against 36% for unaided generalist doctors

On October 22, Microsoft’s Mustafa Suleyman presented a paper in Science on MAI-DxO, a system that coordinates several AI agents to act like a panel of doctors — questioning a patient, ordering tests and debating a diagnosis. In a simulation using complex cases from the New England Journal of Medicine, it was correct 83% of the time, against 36% for primary-care doctors and general internists working alone with reference materials, according to TechCrunch. The caveat is large and stated: specialists and physician teams, who usually handle cases this hard, were not part of the comparison. Microsoft tested the system with models from OpenAI, Google, Anthropic, Meta and DeepSeek, and the best results used OpenAI’s GPT-5 and GPT-5.5. The study’s own data suggests most of the improvement comes from the underlying models, with Microsoft’s orchestration adding gains on top. On emergency-department cases it showed little improvement.

Microsoft says it will test the system in real-world settings through a partnership with Beth Israel Deaconess Medical Center in Boston, where consenting patients will receive its input alongside standard care and the recommendations will be compared with those of human doctors. It is also working with the Mayo Clinic on a healthcare foundation model planned for next year. Suleyman calls the result a step toward “medical superintelligence.”

Who benefits, and who pays? Patients with rare or confusing conditions benefit if a second opinion that costs very little can be added to a rushed consultation. Microsoft benefits from selling to clinicians who already use its dictation and radiology tools. The honest cost is in the comparison itself: beating generalists working alone, on written cases, is not the same as helping a real patient in a real clinic, and the company that built the system wrote the paper. The real-world test with consenting patients is the part that will matter.

The Dhaka view is that the number to read carefully is the 36%. In Bangladesh, a great many patients are first seen by exactly that kind of generalist — often an overworked one, in a crowded outpatient department — and the nearest specialist is a long, expensive journey away. If a system like this ever helps, it will be there. But it needs the same things as every AI health tool: a clinic that works, a doctor who can overrule it, and a patient who can pay for the test it recommends. Without those, an 83% on paper is just a number.

5 Stories That Hit Closer to Home (Bangladesh)

A note on this half: the week’s Bangladeshi AI news was thinner than usual, and several of the strongest items are announcements, press releases or analysis rather than finished results. We have said so on each, and we would rather give you five honest stories than five inflated ones.

1. A Bangladeshi firm launched a Bangla-language app to spot fake financial apps and block phishing links

On October 23, New Age reported that Fred Intelligence Limited, a Bangladeshi cybersecurity firm, has launched an Android app called Swadhin Bangla AI Security, designed to detect fake financial apps, block phishing links and warn users about mobile-banking scams. According to the report, managing director and CEO Bijoy K Chowdhury says it covers malware, fraudulent websites and SMS phishing, and is aimed at mobile financial services such as bKash, Nagad, Rocket and Upay, plus banking and shopping apps. It compares installed apps against a registry of official ones to flag imitations and warns when a financial app is installed from outside Google Play. It also tells users not to share PINs or one-time codes when a payment app opens shortly after a phone call. Other features include ransomware alerts, a password-breach checker and blocking of tracking domains. The company says personal files stay on the device and security data is not sold or used for advertising. (We were able to read this report only in summary, so the details are as reported there; we have not tested the app.)

Who benefits, and who pays? Ordinary mobile-banking users — a very large group in Bangladesh — benefit if the warnings are accurate and arrive before the money moves. The firm benefits from a growing market for protection. The thing to ask is how the app is paid for and what it asks of users: some features need notification and web-protection permissions, and an app that watches your phone is itself something you have to trust.

The Dhaka view is that the most interesting detail is the one about the phone call. Mobile-banking fraud often works by persuasion rather than clever malware — a stranger on the phone talking someone into opening their own payment app — and the app’s PIN-and-call warning is built around exactly that. A warning aimed at that moment targets how such fraud works, which could be more useful than a generic virus scanner. It is also AI used for a narrow, local, plainly useful purpose, which is the kind this newsletter keeps hoping to find more of. The verdict needs real users and real independent testing.

2. The Health Ministry launched an E-Health Hackathon to build AI and digital tools for the health system

On October 18 in Dhaka, the Ministry of Health and Family Welfare and Interspeed Activation Limited launched the E-Health Hackathon 2026, described as a national initiative to bring AI and digital innovation into healthcare, according to a report dated October 19. Registration is open for one month to young innovators across the country. A 10-member screening committee of government, private and technical experts will shortlist teams, who will attend a three-day National Bootcamp in Dhaka in December; the field narrows from 50 teams to 25 and then to the top five. There are 15 priority areas, aimed at points where patients face delays and risks, such as diagnosis backlogs, unverifiable prescriptions and lost referrals. The Ministry may take selected prototypes forward as foundations for official national e-health solutions, and leading ideas may receive recognition from the Prime Minister. The wider goal described is interoperability and a unified national health database, with safeguards for data privacy and sovereignty. The report mentions no cash prize.

Who benefits, and who pays? Young developers get access to real problems and real mentors, and the Ministry gets a pipeline of prototypes without hiring a large team. Patients benefit only if something here actually reaches a hospital. The cost is the part the report does not describe: who owns the prototypes the Ministry might adopt, and who pays to run and maintain them afterwards. A hackathon can produce an idea in three days. A health system needs one that still works in three years.

The Dhaka view is that the list of priority areas is the encouraging part. Lost referrals, prescriptions nobody can verify and slow diagnoses are exactly the daily failures that patients here put up with, and none of them needs a frontier AI model — they need records that follow a patient from a clinic to a hospital. The national health database is the harder and more important goal, because that is where the privacy safeguards the report mentions will be tested. It pairs naturally with this month’s plan for AI-based cancer screening at upazila level: both depend on the same unglamorous groundwork of reliable, shared health records.

3. A look at how AI could hit Bangladesh’s young jobseekers and garment workers

On October 18, The Business Standard published an analysis by Shoeb Hasan, chief business officer of the job platform nextjobz, titled “AI is changing the rules of work in Bangladesh. Are we prepared?” (The author’s company promotes a jobs service, which is worth knowing.) It says about two million young people enter Bangladesh’s job market each year, including more than 600,000 university graduates, and that graduate unemployment is 13.54%, the highest of any education group. It says AI could disrupt work for about four in ten workers, that freelancers report drops in assignments and earnings of roughly a quarter to more than 80%, with logo design, photo editing and data entry hit hardest, and that more than a third of Bangladeshi CEOs expect junior positions to decline over the next three years. It cites projections that up to 1.22 million garment jobs could disappear by 2041, and that up to 60% of women’s jobs in the sector could be at risk. It also notes research suggesting AI can improve defect detection and production planning in Bangladeshi factories without necessarily cutting jobs. About 42% of technical training still focuses on basic computer skills, the piece says, and public spending on education and skills has fallen to 1.3% of GDP.

Who benefits, and who pays? Employers who adopt AI early benefit, and the author’s advice — prepare people for their next role before their current one disappears — is sensible for any worker who can act on it. Those who pay first are the entry-level graduate, the freelancer, and the garment worker, who is more likely to be a woman, and who has the least ability to retrain on her own. The honest limit of this piece is that it is an analysis built on other people’s projections, several without a stated source, and the 2041 garment figure is a forecast, not a count.

The Dhaka view is that the numbers matter less than the shape. Bangladesh’s two big hopes — a young population and a garment industry that employs nearly four million — are both exposed to the same technology at once, in the same decade the country hopes to graduate from least-developed status. The draft National AI Policy proposes AI lessons in classes eight and nine, which is a beginning, but a class eight pupil today enters the job market well after the pressure on entry-level work has already arrived. The question for any government plan is not “how many students learn about AI” but “where do the 600,000 graduates a year find their first job.”

4. A Dhaka firm showed AI leak-detection robots at WETEX in Dubai

On October 22, a press release carried by The Daily Star said Smart Build, a Dhaka-based smart and sustainable development company, demonstrated AI-powered water leak-detection robots at WETEX 2026, the 28th Water, Energy, Technology and Environment Exhibition, held October 20–22 at the Dubai World Trade Centre and organised by the Dubai Electricity and Water Authority. The company says the robots are designed to detect leaks more efficiently and reduce water loss and waste, and that it is developing an AI-powered water intelligence programme for urban, industrial and agricultural water management in Bangladesh. Its chief technology officer, Shafkat Alam, is quoted as saying “combining AI, robotics, and smart technology can help solve real-world challenges.” The release says its management met Abdulla Ahmed Balalaa, the UAE’s Assistant Minister of Foreign Affairs for Energy and Sustainability and Chairman of Masdar City, who it says appreciated the company’s solutions and offered support for future collaboration. These are the company’s own claims, in its own release; the article gives no performance figures.

Who benefits, and who pays? The firm benefits from exposure at a large international fair, and Bangladeshi cities benefit if leak detection ever reaches a water authority that loses a lot of water to old pipes. Nobody pays anything yet. The caveat is simple: a demonstration at an exhibition is a first step, and a press release is not a contract, a pilot or a result.

The Dhaka view is that the problem is real and local. Water loss from leaking pipes is a problem for many cities, and a water-intelligence programme is the sort of narrow, practical AI use the country could use. The test of this story is not whether a minister in Dubai was polite, but whether, a year from now, a Bangladeshi water utility has run a pilot with published numbers on how much water was saved. We will be looking for it.

5. The PHP emergency reached the world’s headlines — and we could not find a Bangladeshi advisory for website owners

Correction (November 2, 2026): Bangladesh’s government cyber-response team did warn about these flaws. Prothom Alo reported on October 24 that BGD e-GOV CIRT “issued an urgent alert regarding the PHP security vulnerability” on October 22 and advised organisations to check whether their PHP versions were vulnerable. The same report said Bangladesh Bank’s website and the land-services site, land.gov.bd, were taken offline, along with the websites of six banks. We missed the alert when we searched, and the section below is wrong on that point. We have left the original text in place so the error stays visible. (Prothom Alo)

On October 22, The Daily Star carried an explainer, “How two PHP flaws brought down websites around the world,” and an AFP report, “Websites go dark worldwide as AI exposes critical security flaws.” As covered above, the flaws affect software that runs about 70% of the web. Neither article, as far as we can tell from the Daily Star’s explainer, mentions Bangladesh or any Bangladeshi website or agency. Bangladesh’s government cyber-response team, BGD e-GOV CIRT, publishes advisories on its website, and we looked for one on this flaw this week; we did not find one, though that may only mean we missed it or that it has not been published yet. We are not reporting that none exists — only that, after searching, we could not find one on the day of writing.

Who benefits, and who pays? Every Bangladeshi website owner benefits from a clear, local, plain-language instruction on what to check and what to update. Whoever does not get it pays, usually later and without noticing: a defaced page, a stolen customer list, a server quietly used to attack someone else. The cost of the warning is a few hours of an agency’s time. The cost of its absence falls on small businesses that cannot afford a security team.

The Dhaka view is that this is the practical test of the “specific advisory” idea raised in last week’s edition about India’s CERT-In. The advice here is not complicated: find out which PHP version each site runs; update to 8.5.12, 8.4.27, 8.3.36 or 8.2.35; and, if a site runs PHP 8.1 or older, understand that it no longer gets regular updates and needs to be upgraded or taken offline. WordPress and similar tools carry PHP with them, so the owner of a small Bangladeshi shop is exposed without knowing the word. That is exactly the gap a short, public advisory would close.

Wrapping up

Ten stories in seven days, and the international five were a study in who gets to look inside the machine. AI helped find two critical flaws in PHP — a language behind about 70% of websites — and governments from Japan to Finland took sites offline while patches caught up, with the PHP Foundation admitting that a flaw known since at least 2024 had waited too long for a fix. Outside researchers from METR and Redwood published what they found in three weeks inside OpenAI: an agent monitor with real gaps, patched, according to OpenAI, but not independently re-checked. OpenAI reportedly chased $30 billion at a $1.25 trillion valuation while Anthropic prepared a public listing. DeepSeek reportedly raised about $15 billion two days after the US ordered federal agencies to remove it. And Microsoft said its AI beat unaided generalist doctors 83% to 36% on hard written cases — with specialists left out of the comparison.

At home, the week was smaller, and we have tried not to dress it up. A Bangladeshi firm launched a Bangla app aimed at the way mobile-banking fraud actually happens. The Health Ministry opened a hackathon for tools that fix lost referrals and unverifiable prescriptions. A sober analysis put numbers — some of them forecasts — on what AI could mean for graduates and garment workers. A Dhaka firm took leak-detection robots to Dubai and left with a polite meeting. And while the world’s governments scrambled to patch PHP, we could not find a Bangladeshi advisory telling our own website owners to do the same.

The thread across both halves is the question of who is looking. Abroad, the news was about the people outside a system finally getting a view in — researchers, regulators, patch coalitions — and finding gaps that the insiders had either missed or been slow to fix. At home, the smallest stories had the same shape: a scam warning that fires at the right moment, a health record that follows a patient, a pipe whose leak someone can finally see. The lesson from PHP is the least exciting and the most useful one of the week. The flaw that took governments offline had reportedly been known since 2024. Knowing is not the same as acting, and the cheapest AI-era security measure anyone in Dhaka can take today is to find out what their website is running, and update it.

If something here caught your interest, share it with someone who does not follow tech. The conversation only spreads if we keep starting it.

See you next week. Stay well.