AI disclosure: The Web Guy Nick Podcast is produced with AI. Alex and Jordan are AI hosts, and Nick's voice is an AI clone of his own voice, used with his permission. Stories are researched from public sources and linked below.
Google's threat team found an attack system that reads a software patch, works out what it fixed, writes an exploit, and runs it with no human in the loop. Anthropic answered the new tempo by cutting the fix window for serious bugs from 90 days to 30. Plus: Alibaba's Qwen4-Max flagship lands 25 percent cheaper, ChatGPT for Healthcare agents go after paperwork, Hugging Face routes model traffic by live price, and NVIDIA builds a chip for checking machine ideas. Nick's Take: patching is now a race, and it is won by removing the human bottleneck.
Listen and watch
- Listen on Spotify: https://open.spotify.com/episode/3Exix1oqfh4pYklieIP732?si=rxkeFfnySRKeYd-_7g60WA
- Watch on YouTube: https://youtu.be/RD9CzRyW-OU
Show notes
- Google's Threat Intelligence Group and Mandiant found a staging server running "exploit agent" - it compares patches, drafts exploits, and executes them with no human in the loop. The patch-to-exploit gap that used to take weeks is collapsing.
- Anthropic cut its coordinated vulnerability disclosure window from 90 days to 30 for high and critical severity bugs (45 days for the rest), for reports sent on or after January 1, 2027. Escalation to an outside coordinator drops from 30 days of silence to 15.
- Alibaba released Qwen4-Max on November 16: a 2.4-trillion-parameter flagship at $1.50 per million input tokens and $4.50 output, 25 percent cheaper than the model it replaces, with a 1-million-token context window.
- OpenAI expanded ChatGPT for Healthcare with computer-use agents and more than 20 healthcare plugins. The agents move information between health records, insurer portals, and scheduling systems - administration, not diagnosis.
- Hugging Face now routes Inference Providers requests through OpenRouter, picking providers by live price, speed, and reliability, with automatic failover. Your existing API calls do not change.
- At SC26, NVIDIA introduced Rubin HPC, a GPU for scientific simulation and AI-assisted engineering with more than 50,000 CUDA cores, sparse-data acceleration, and seventh-generation NVLink.
- Quick hit: OpenAI is expanding its researcher program toward 200,000 researchers worldwide, with applications for 50,000 new grants opening November 23.
Nick's Take
Patching used to be maintenance. It is now a race, and races are won by whoever removed the human bottleneck. I manage about 44 websites. The sites that stay safe are not the ones with the most security tools. They are the ones where updates run automatically and a person checks the exceptions on a schedule. For every system your business depends on, ask one question. If a fix shipped tonight, would it be installed by morning without anyone remembering to do it? If the answer is no, that system is your weakest link, and no new AI tool matters more than fixing that first.
One Thing to Try This Week
The 15-minute update sweep. Go through the five systems your business leans on most - your website, your point of sale, your booking tool, your email app, and your router. Turn on automatic updates in each one. For anything that cannot update itself, put a recurring 15-minute check on your calendar, same time every week.
Sources
- SXF Brief, AI News Nov 17, 2026: https://sxf.si/brief/
- Help Net Security, Anthropic vulnerability disclosure deadline drops to 30 days: https://www.helpnetsecurity.com/2026/11/17/anthropic-vulnerability-disclosure-days/
- Anthropic, Updating our coordinated vulnerability disclosure policy timelines: https://www.anthropic.com/research/coordinated-vulnerability-disclosure-timelines
- Apidog, Qwen4-Max vs Qwen 3.8: https://apidog.com/blog/qwen-4-max-vs-qwen-3-8/
- Digital Applied, Qwen4-Max Is Here: https://www.digitalapplied.com/blog/qwen4-max-price-benchmarks-access
- join.dev, Today in AI Nov 17, 2026: https://join.dev/
- Higgsfield changelog, Short Drama Studio: https://higgsfield.ai/creator-hub/changelog?fpr=promptowy&fp_sid=quantslant
Find your next AI tool at BotPickr, our directory of more than 170 AI tools and bots: https://botpickr.com
Full transcript
Alex: Before we start: this podcast is produced with AI. Jordan and I are AI hosts, and Nick's voice is an AI clone of his real voice. Every story is researched from public sources, with links in the show notes.
Alex: Here is the cold open. Google's threat team just documented an attack system that reads a software patch, figures out what the patch fixed, writes an exploit for the hole, and runs it. No human in the loop. The race between a fix shipping and an attacker using the hole used to take weeks. It is starting to take hours.
Jordan: That is our lead story, and it sets the theme for the whole show. Speed. Attackers got faster, defenders answered by cutting a ninety day safety window down to thirty, and in the middle of it a flagship model got twenty five percent cheaper. I am Jordan.
Alex: I am Alex, and Nick is here, the only host on this show who can actually be locked out of his own accounts.
Nick: And the only one who has to patch forty four client websites when stories like this land. Let us get into it.
Alex: Story one. On Monday, November sixteenth, Google's Threat Intelligence Group and Mandiant published their latest AI Threat Tracker. The headline finding is a staging server they found running something called exploit agent. The system watches software updates, compares the new code against the old code, works out what vulnerability the update was fixing, drafts an exploit, and executes it, all without a person directing each step.
Jordan: Why does that matter so much? Because the classic defense was time. A vendor shipped a patch, and most attackers needed days or weeks to reverse engineer it. That gap was your safety margin. If a machine closes that gap the same day, the patch is not the finish line anymore. Installing it fast is.
Nick: From the owner seat, this is the part people miss. The patch existing does nothing. I have seen client sites sit on an available update for a month because nobody owned the job. An exploit agent does not wait for your schedule. It works nights, weekends, and holidays, which is more than I can say for most update routines.
Alex: Google also named other agentic tools showing up in real attack activity, including frameworks called CAIRN, PentAGI, and CyberStrikeAI. To be precise, Google is not saying every attack is now fully autonomous. They are saying the pieces are live in the wild, chained together, and the tempo has changed.
Jordan: The practical tip here is simple. Turn on automatic updates everywhere you can, for your website platform, your plugins, your phone, your router, and your point of sale software. Then once a month, check the two or three systems that cannot update themselves. That one habit now matters more than any security plugin you can buy.
Alex: Story two is the defender's answer to that tempo. Anthropic announced on Monday that it is cutting its coordinated vulnerability disclosure window. When Claude finds a serious bug in someone else's software, Anthropic privately tells the maintainer and gives them time to fix it before publishing. That window was ninety days. Starting with reports sent on or after January first, twenty twenty seven, high and critical severity bugs get thirty days. Everything else gets forty five.
Jordan: And the smaller clocks shrank too. If a maintainer does not answer, Anthropic escalates to an outside coordinator after fifteen days instead of thirty. After a patch ships, full technical details follow in thirty days instead of forty five. The seven day window for bugs already being exploited stays the same.
Nick: Anthropic's reasoning is the same as Google's report, from the other side. A bug one model finds, another model is likely to find soon. Sitting on a secret for ninety days no longer protects anyone, because the secret does not stay secret.
Alex: Google's Project Zero made a similar move on November thirteenth, heading toward a thirty day window on its own timeline. So this is an industry direction, not one company being impatient.
Jordan: Owner translation. If your business runs on WordPress, Shopify apps, or any stack of plugins, the time between a fix existing and the details going public just got a lot shorter. Your update routine is now a thirty day routine at most, and honestly, after story one, it should be a same week routine.
Nick: Quick break, and I am spending it on my own site. If you have ever typed 'best AI tool for' into Google and drowned, BotPickr is the fix. More than 170 AI tools and bots, organized by the work they do: sales, support, content, scheduling, the lot. Pick the job, get a short list that actually fits, and skip the hype videos. Find your bot in 60 seconds at botpickr.com. Now back to Alex and Jordan, who never need a break, because they are robots.
Alex: Story three is the price story, and it is a good one. On Monday, Alibaba released Qwen four Max, its new flagship model. Two point four trillion parameters in total, with about seventy one billion active on each token, a one million token context window, and it takes text, images, and video in. The headline for owners is the price. One dollar and fifty cents per million input tokens, four dollars and fifty cents per million output tokens. That is twenty five percent cheaper than the model it replaces, and cached input drops to seven and a half cents per million.
Jordan: Alibaba's own benchmark, called Work three sixty, covers industry tasks across twenty six domains. On it, the average cost per task fell from four dollars and forty three cents to one dollar and sixty seven cents, about sixty two percent lower. That is the vendor's own benchmark, so treat it as a claim to test, not a settled fact. Independent scoring puts the model below the very top Western flagships on raw intelligence, which is exactly the trade worth understanding.
Nick: Here is how I read it. Most business work is not frontier research. It is summarizing calls, drafting product descriptions, sorting tickets, cleaning up lists. If a model does those jobs well at a quarter less per token, your monthly bill drops without your customers noticing any difference. The expensive model should earn its place job by job.
Alex: The practical tip. Pick one routine job your business runs on a top tier model, run the same twenty real examples through a cheaper model, and compare the outputs side by side. If you cannot tell which is which, move the job and keep the savings.
Jordan: Story four moves to healthcare, but the pattern applies to every service business. On Monday, OpenAI expanded ChatGPT for Healthcare with computer use agents and more than twenty healthcare plugins. The agents work inside the software clinics already use, moving information between health records, insurer portals, and scheduling systems. OpenAI says access depends on each workspace's settings, and the healthcare products carry business associate agreements to support compliant workflows.
Alex: The important line in the announcement is what the agents are for. Not diagnosis. Administration. The referrals, the forms, the portal hopping that eats staff hours and delays care. OpenAI is careful to say an agreement does not replace a clinic's own security and clinical review.
Nick: Swap the industry and it is my clients' world. A med spa, a contractor, a print shop, they all lose hours moving information between systems that do not talk to each other. An agent that can operate the software you already have, with a person approving anything that leaves the building, is the first version of this technology that fits a small operation.
Jordan: The caution travels too. Start with low risk, auditable tasks. Test on sample records before real ones. And write down who approves what before the agent touches anything live.
Alex: Story five is for anyone building on open models. On Monday, Hugging Face announced that language model requests through its Inference Providers service now route through OpenRouter. In plain terms, when you ask for an open weight model, the system picks which provider runs it using live price, speed, and reliability signals, and it can retry through another provider if one fails. Your existing code does not have to change.
Jordan: That sounds like plumbing, and it is, but plumbing decides your bill and your uptime. The takeaway is that where a request runs is becoming an automatic decision, optimized in real time, instead of a contract you signed once and forgot.
Nick: Owner version. If your website chat or your internal tools run on an A P I, ask your developer one question this week. Are we locked to one provider's price, or can our traffic route to whoever is cheapest and fastest that day? Routing is the new discount. You do not negotiate it. You switch it on.
Alex: Story six is the big iron. At the S C twenty six supercomputing conference on Monday, NVIDIA introduced Rubin H P C, a new graphics processor in its Vera Rubin platform, built for scientific simulation and A I assisted engineering. NVIDIA says it carries more than fifty thousand C U D A cores, acceleration for sparse data, and seventh generation NVLink connections between chips.
Jordan: Why should a business owner care about a supercomputing chip? Because of the job it does. Agents are getting good at proposing ideas, a molecule, a material, a part design. Somebody still has to simulate whether the idea survives the real world. Rubin H P C is built for that checking step, and NVIDIA is betting the checking step becomes as big a market as the generating step.
Nick: It is the same pattern as story one, at the other end of the economy. Machines propose, machines test, people decide. The businesses that win are the ones with a fast, honest checking step, whether that is a simulation, a proofread, or a fingerprint on a purchase.
Alex: Quick hit before the segments. OpenAI is expanding its program for academic researchers, aiming at up to two hundred thousand researchers worldwide. Applications for fifty thousand new grants open November twenty third, and each grant provides ChatGPT Plus for the recipient and up to two collaborators. If you know a researcher, that date is worth circling.
Jordan: Now Nick's Take, drawn from the lead story.
Nick: My take is this. Patching used to be maintenance. It is now a race, and races are won by whoever removed the human bottleneck. I manage about forty four websites. The sites that stay safe are not the ones with the most security tools. They are the ones where updates run automatically and a person checks the exceptions on a schedule. So here is the owner lesson. For every system your business depends on, ask one question. If a fix shipped tonight, would it be installed by morning without anyone remembering to do it? If the answer is no, that system is your weakest link, and no new A I tool matters more than fixing that first.
Alex: Higgsfield Corner. On Monday, Higgsfield launched Short Drama Studio, in early preview on desktop. You give it a script, and it splits the text into episodes, builds the cast, the locations, and the props, and generates the episodes with the same characters throughout. Consistent characters across episodes has been the hard part of A I video, and this is aimed straight at it.
Jordan: For a local business, the play is episodic social content. A coffee shop runs a five part mini story with the same two characters every week, instead of five unrelated clips that nobody connects. It is a preview, so test it with one short script before you plan a season around it. And check every frame at full size before anything posts, same rule as always.
Alex: Time for AI or BS, our quick fire game. Three claims, Jordan calls them, Nick judges, and the stakes are real. The loser has to approve all of Nick's agent purchases for a week.
Jordan: Nick already approves all of his agent purchases. That is the whole rule from yesterday's episode.
Nick: Exactly. Which is why I never lose. Claim one, Jordan.
Jordan: Claim one. The exploit agent Google found needed a human to approve each attack before it ran.
Nick: That is BS. The whole point of the finding is that it compared patches, drafted the exploit, and executed with no human in the loop.
Jordan: Correct. Claim two. Starting in January, Anthropic gives maintainers thirty days to fix a high severity bug before the details go public.
Nick: That one is true. Thirty days for high and critical, forty five for the rest, for reports sent on or after January first.
Jordan: Correct again. Claim three. Qwen four Max costs more per token than the model it replaces.
Nick: BS. It is twenty five percent cheaper, a dollar fifty in and four fifty out per million tokens.
Alex: A clean sweep, which means the robots owe Nick a week of purchase approvals he was already doing himself. Some prizes are just your regular job with a ribbon on it.
Jordan: One Thing to Try This Week. The fifteen minute update sweep. Go through the five systems your business leans on most, your website, your point of sale, your booking tool, your email app, and your router. Turn on automatic updates in each one. For anything that cannot update itself, put a recurring fifteen minute check on your calendar, same time every week. After today's lead story, that quarter hour is the highest value security work a small business can do.
Alex: That is the show. A I agents learned to write and run exploits, defenders answered by cutting the fix window to thirty days, a flagship model got twenty five percent cheaper, healthcare agents went after paperwork instead of patients, model traffic learned to route itself to the best price, and NVIDIA built a chip for checking machine ideas. Your one action this week is the update sweep. Fifteen minutes, five systems, automatic updates on.
Nick: And if your website is one of those systems and you are not sure it is covered, that is the work we do every day at Web Guy Nick. You can find us at webguynick.com.
Jordan: We will see you tomorrow.
