What's New in Tech

🎙️ GPT-6 Cyber Finds Thousands of Vulnerabilities in One Run, Taiwan Stocks Close Down 4.5% the Same Day, What Happened | AI Daily Podcast

Mark Ku

November 5, 2026•23 min read

View article sources

Opening

OpenAI today unveiled GPT-6 Cyber, a security-focused model that, according to reports, surfaced over a thousand vulnerabilities in a single automated run, burning through roughly $1.2 m i l l i o n i n A P I c r e d i t s a l o n g t h e w a y . M e a n w h i l e, m i d t e r m e l e c t i o n r e s u l t s s e n t t e c h s t o c k s i n t o t h e r e d a c r o s s t h e b o a r d, w i t h T a i w a n^{'} s T A I E X c l o s i n g d o w n 4.5 1.2 million in API credits along the way. Meanwhile, midterm election results sent tech stocks into the red across the board, with Taiwan's TAIEX closing down 4.5% on the same day, memory and AI supply chains taking the brunt. And as for why Apple started cutting its engineers' daily AI budgets right after posting$ 114.5 billion in quarterly revenue, I'll get to that shortly.

Today's Top Stories

1. OpenAI launches GPT-6 Cyber: AI uncovers 81 Firefox vulnerabilities and 948 Linux privilege escalation bugs

Source: The Hacker News (https://thehackernews.com/2026/11/openai-launches-gpt-6-cyber-after-ai.html)
Summary: OpenAI released GPT-6 Cyber, a security-oriented model whose headline feature is automated vulnerability discovery. Per The Hacker News on November 5, the pipeline found 948 Linux local privilege escalation vulnerabilities, 81 high-risk Firefox vulnerabilities, and one critical Brotli bug rated CVSS 9.2. The report says the whole experiment ran about 30 hours and consumed roughly $1.2 m i l l i o n i n A P I c r e d i t s, a v e r a g i n g a b o u t$ 1,279 per vulnerability.
Most surprising takeaway: For the first time, vulnerabilities have become something you can price by dividing an API bill by a unit count, and the unit price is low enough to make human bounty hunters wince.
Taiwan angle: Firmware at Taiwan's many networking, IoT, and industrial control vendors has long relied on manual audits and external pentests. If both attackers and defenders can scan at this cost, Taiwanese manufacturers will have to completely rethink their pre-shipment security validation cadence.
Discussion points:
1. When the cost of finding a vulnerability drops to four figures, can bug bounty economics still hold up?
2. If the same capability lands in attackers' hands, can defenders patch fast enough to keep pace?
3. Most open source maintainers are volunteers. Who's going to triage hundreds of vulnerability reports at once?
Script suggestions: Let me be upfront: all of today's numbers come from the reporting, and we haven't independently verified them. But this one is genuinely vivid, thirty hours, $1.2 m i l l i o n, o v e r a t h o u s a n d v u l n e r a b i l i t i e s . A q u i c k a r i t h m e t i c c a v e a t :$ 1.2 million divided by the total vulnerability count doesn't equal $1,279, so that per-unit figure must come from some specific calculation the report didn't spell out. What I care about more is the follow-on question: open source maintainer capacity hasn't grown, so what happens to disclosure and patching timelines? Nothing in what I've seen explains the disclosure schedule, and that's worth watching.

2. Midterm "anti-AI wave" hits data center stocks: US markets tumble, TAIEX closes down 4.5%

Source: NBC News (https://www.nbcnews.com/business/markets/tech-stocks-plunge-ai-midterms-rcna607650)
Summary: Per NBC News on November 5, the midterms produced a cohort of winners running on data center restrictions, and markets immediately read it as policy risk for AI infrastructure. The Nasdaq fell over 5% that day, and Taiwan's TAIEX closed down 4.5% the same day, with memory and AI supply chain names hit hardest. A note on timing: the report refers to the selloff on "Thursday," and November 5 happened to be a Thursday.
Most surprising takeaway: What tapped the brakes on the AI trade wasn't earnings or chip yields, it was a batch of local ballots.
Taiwan angle: Taiwan's market wasn't just collateral damage this time; it was sold as a directly correlated proxy for AI capex. The amplified volatility in memory names means the market now sees Taiwanese order books as tied directly to when US data centers break ground.
Discussion points:
1. NBC's headline used "plunge," but there's usually a gap between campaign promises and actual regulation. Is this a sentiment correction or a structural turn?
2. If data center reviews become routine for local governments, what metrics should we use to track order visibility for Taiwanese suppliers, grid interconnection approvals, land development permits, or customer capex guidance?
3. Could local issues like electricity, water, and noise push the next wave of data center siting offshore? Would Taiwan have a shot at winning any of it?
Script suggestions: The thing I think Taiwanese listeners should take away is that our index may now be more sensitive to US local politics than to our own fundamentals. We used to track capex from the big four North American cloud providers; now we probably need an extra column for whether local councils will even let data centers break ground. For investors, that means more sources of risk. For equipment and memory makers, orders may not shrink, but lead times and revenue recognition get stretched, and share prices tend to price in the latter first.

3. Isomorphic Labs reportedly in talks for $5 B r a i s e a t a$ 50B valuation, with no drug in human trials

Source: The Next Web (https://thenextweb.com/news/isomorphic-labs-5bn-funding-talks-50bn-valuation)
Summary: Per The Next Web on November 5, citing Bloomberg, DeepMind spinout Isomorphic Labs is in talks for a new round of roughly $5 b i l l i o n a t a v a l u a t i o n r e a c h i n g$ 50 billion. That's about six months after the $2.1 billion round the report mentions from May. The key detail: so far, it has no drug in human clinical trials.
Most surprising takeaway: A company that hasn't put a drug into a human body is approaching the valuation of established pharma companies. What the market is buying is pure AlphaFold pedigree and platform narrative.
Taiwan angle: Taiwan's biotech sector has long been stuck at "great technology, trials too expensive." The valuation logic behind these AI drug discovery platforms is a reminder that the language of fundraising has changed entirely.
Discussion points:
1. What metric should actually measure the value of AI drug discovery, number of candidate molecules, number of pharma partnerships, or eventual clinical success rate?
2. A valuation jumping a whole tier in six months: overheated capital, or genuine milestone progress?
3. Can Taiwan's CROs, CDMOs, and protein structure research groups plug into this supply chain?
Script suggestions: I'm optimistic about AI in drug discovery, but let me temper things a bit. The hardest part of drug development was never finding the molecule, it's toxicity, dosing, and the pile of human-body reactions you can't predict, all of which cost time and real money. So what this round is buying is the ability to compress the front end, not a guarantee of skipping the back end. For Taiwan, the interesting angle is a different one: if the front end mass-produces candidate molecules, demand for downstream synthesis, contract manufacturing, and animal studies all scales up, and we actually have a foundation there.

4. Apple posts $114.5B in quarterly revenue, then cuts some engineers' daily AI spend cap by a third

Source: Inc. (https://www.inc.com/moses-jeanfrancois/apple-posted-114-billion-dollars-revenue-now-cutting-some-engineers-daily-spending-limits/91431562)
Summary: Per Inc. on November 5, citing Bloomberg, Apple in October lowered the daily AI usage cap for some engineers from $300 t o$ 200, with some teams possibly dropping as low as $25. I n t h e s a m e q u a r t e r, A p p l e r e p o r t e d$ 114.5 billion in revenue, up roughly 12% year over year. The report also notes the tightening has stirred layoff anxiety internally.
Most surprising takeaway: A company doing $100B+ in quarterly revenue is installing electricity meters on its employees' tokens.
Taiwan angle: Taiwanese companies adopting AI often get stuck at "just buy the credits first." Apple's move previews the next phase of the problem: not whether you can use it, but how to measure how much output each dollar of tokens buys.
Discussion points:
1. Does a $25 daily cap amount to a throttle for engineers who lean heavily on coding agents?
2. Once AI costs shift from experimental budget to recurring expense, what KPI proves it's worth it?
3. How should a company communicate usage controls so it doesn't fuel layoff anxiety and shoot itself in the foot?
Script suggestions: The signal I read here is that AI has dropped from the "strategic investment" bucket into the "operating cost" bucket. Nobody questions your spend when it's a strategic investment; once it's an operating cost, the CFO pulls out a calculator. A lot of Taiwanese companies are one stop behind Apple right now, still talking about adoption, still talking about subscriptions. But starting next quarter, they'll face the same question: how much faster did this AI spend actually make delivery? My own advice: rather than slashing budgets outright, first build usage logs mapped against output. Otherwise the money you save gets paid for with your engineers' time.

5. Anthropic signs its first content licensing deal, with the Associated Press

Source: Dataconomy (https://dataconomy.com/2026/11/05/anthropic-signs-first-content-licensing-deal-with/)
Summary: Per Dataconomy on November 5, Anthropic signed its first content and data licensing agreement with the Associated Press, with neither the amount nor the term disclosed. In practice, the deal lets Claude display real-time election returns and polling data with AP source attribution on midterm election night. The report highlights the reversal: Anthropic has long argued that scraping content falls under fair use.
Most surprising takeaway: What pushed a fair-use hardliner into paying wasn't litigation, it was that election night data can't wait.
Taiwan angle: This is a reality check for Taiwanese media: what models will actually pay for is real-time, structured, attributable data. A pure archive of articles has comparatively little bargaining leverage.
Discussion points:
1. Will licensing become the exclusive domain of large news organizations, making small outlets even harder to find?
2. Real-time data is worth far more than historical content. How should Taiwanese media restructure their data assets?
3. Does source attribution actually increase user trust in AI answers?
Script suggestions: I think this is more interesting than it looks on the surface. The debate over the past two years has been stuck on whether training data counts as fair use, but what got broken through here is a different line: immediacy. You can argue archived articles were scraped long ago, but tonight's vote counts? Without a license, you simply can't get them, and one wrong number is a PR disaster. For Taiwanese media, the real question isn't "how do I stop AI from scraping me," it's "do I have a clean, real-time, licensable data product?" If you do, you have a seat at the negotiating table. If you don't, you just keep watching other people sign deals.

Closing

These five stories actually thread together: the cost of AI, its political risk, and the price of data are all being repriced in the same quarter, from GPT-6 Cyber's per-vulnerability price tag, to Apple's daily caps, to Anthropic finally paying for real-time news. One more reminder: all the figures above come from the reporting and we haven't independently verified them, so take them with appropriate caution. I'm Mu-Yan, thanks for listening to today's Mark's Tech Insights. See you next episode.