It's Happening... AI Just Started the Biggest Internet Nightmare in History — Transcript
Full transcript
- 0:0227 major cyber attacks and security
- 0:04incidents hit the internet in just 48
- 0:07hours. Governments across Europe yanked
- 0:09their own websites offline. Japan's
- 0:12prime minister went on X asking regular
- 0:14people to check their servers. And a
- 0:16company that vouches for websites being
- 0:18safe turned out to be guarding its
- 0:20master keys with the number four. People
- 0:22are calling it the darkest October and
- 0:25AI is in the middle of almost all of it.
- 0:27finding the holes, exploiting them, and
- 0:29sometimes wrecking things all by itself.
- 0:32So, every app sits on a stack of
- 0:34software layers. And this week, nearly
- 0:36every layer cracked. [music] At the
- 0:38bottom is Linux, running most of the
- 0:41world's servers and basically the whole
- 0:43cloud. Cloud companies split one big
- 0:45machine into sealed off virtual machines
- 0:48rented to different customers. On
- 0:50October 15th, a bug went public that let
- 0:53someone inside one of those break out
- 0:55and grab root, meaning full admin
- 0:57control over the host and everyone on
- 0:59it. An autonomous AI security engine
- 1:02from pone.ai found it, and Lionus
- 1:05Torvalds merged the fix a day earlier.
- 1:07Add another Linux bug handing regular
- 1:10users admin rights, plus a TIDB database
- 1:13flaw. One layer up is EngineX, the
- 1:16software standing at the front door of
- 1:18millions of websites. On October 14th,
- 1:20it patched five bugs, three rated 9.2
- 1:23out of 10. Some let a crafted request
- 1:26overflow memory and run the attacker's
- 1:28code or crash the site. And one breaks
- 1:30the check where both sides show ID
- 1:32badges so attackers walk right past it.
- 1:34Then GitLab, where tons of companies
- 1:37keep their code and the automated
- 1:39pipelines that push updates live. It got
- 1:42two perfect 10 out of 10 bugs on October
- 1:4412th and 21st, letting anyone take over
- 1:47a self-hosted GitLab server without
- 1:49logging in. On October 20th, Chrome 156
- 1:52shipped with 353 security fixes. A year
- 1:56ago, Chrome 141 and 142 had 21 and 26.
- 2:01Six were critical, including a 9.6 where
- 2:04a booby trapped web page could escape
- 2:06Chrome's sandbox. Four of those six came
- 2:09from anthropic researcher Xinyang J, one
- 2:12found with Claude's help, and one more
- 2:13from OpenAI's codec security. So, five
- 2:16of six came from AI companies. Then came
- 2:19the night of October 21st. Finland shut
- 2:22down all its government's online
- 2:23services. The Netherlands took down its
- 2:26digital ID system, Digid, government
- 2:28portals, the Port of Roderdam, and
- 2:30several hospitals. And even the Dutch
- 2:32cyber security center switched off its
- 2:34own site, which is a bit like the fire
- 2:36department evacuating the fire station,
- 2:38but honestly, correct call. France
- 2:41ordered every provincial government site
- 2:43offline, and Germany, Belgium, Wikipdia,
- 2:46and Nintendo pulled services, too. In
- 2:48Japan, Pixive and Nikono went into
- 2:51emergency maintenance. The sushi chain
- 2:53Sushiro paused its ordering app, and
- 2:55Final Fantasy 14 players in China had to
- 2:57switch launchers. All of them unplugged
- 3:00themselves before anyone broke in. The
- 3:02reason is PHP, the language running the
- 3:04back kitchen of the internet. Log in,
- 3:07search or comment on a site, and PHP is
- 3:10usually what's handling it. WordPress,
- 3:12Drupal, Joomla, Magento, and the
- 3:15software behind Wikipedia all run on it.
- 3:17And W3 Tech says it powers 69.3% of
- 3:21websites with an identifiable backend
- 3:23over 33 million sites, while second
- 3:26place sits at 8.1%.
- 3:29The two bugs, nicknamed PHP2 Shell, let
- 3:32one request get an attacker a shell, a
- 3:35command line on your server, so total
- 3:37control, no password needed. Both are
- 3:39rated 9.2 and Watchtower gave them a
- 3:42full 10. The first one is about running
- 3:44out of memory. PHP slams an emergency
- 3:48stop and at exactly the wrong moment
- 3:50that leaves data half set up which the
- 3:52cleanup treats as real. So on Apache an
- 3:54attacker can run code before the
- 3:56website's own code even starts. The
- 3:58second is about leftovers where the next
- 4:01request reuses memory PHP already threw
- 4:04away which can run code or leak other
- 4:06users passwords and API keys. Whiz said
- 4:09even a bare hello world site is
- 4:11affected. Cloudflare says this goes back
- 4:13over a decade to PHP 5.4. 4 and Vulkar
- 4:17Dush of the PHP Foundation compared it
- 4:19to heartbleleed in 2014. Patches only
- 4:22cover versions 8.2 through 8.5, while
- 4:2527% of PHP sites still run version 7 and
- 4:297.7% run version 5, which get no fix.
- 4:32The release was planned for October
- 4:3422nd, but on the 20th it leaked on
- 4:36Hacker News and Breach forums.
- 4:38Watchtower confirmed it within hours and
- 4:40PHP moved the patch up a day. That
- 4:43night, Califf, one of the finders,
- 4:45accidentally posted a demo of its attack
- 4:47hitting real targets. And next morning,
- 4:49Chain Guard, a distributor with early
- 4:51access, broke its non-disclosure
- 4:53agreement and shipped the fix early,
- 4:55which is basically a treasure map for
- 4:57attackers. Cloudflare saw attacks 6
- 5:00hours before the official patch. The
- 5:02NSA, Britain, and the EU said to take
- 5:04unpatched sites offline, and Prime
- 5:07Minister Sanai Takahichi told Japan to
- 5:09get ready for the 11 p.m. patch. OpenAI
- 5:12then counted PHP2 Shell as an AI win for
- 5:15GPT 5.6 Soul, its cheaper model. Califf
- 5:19says researcher Ryan Austin found it
- 5:21first on August 5th using public models,
- 5:24and his edge was history. Stefan Esser
- 5:26found a similar PHP memory bug in 2004.
- 5:29[music]
- 5:30A related 2020 bug report got ignored,
- 5:32and a 2024 audit funded by Germany's
- 5:35sovereign tech agency spotted this exact
- 5:37flaw. But nobody could prove it was
- 5:39exploitable and every fix broke
- 5:41something so it stalled. Writer Nick
- 5:43Rogers calls this security by
- 5:44exhaustion. Turning a bug like this into
- 5:46a working attack took weeks of tedious
- 5:49manual debugging so nobody bothered and
- 5:51tons of old internet code is protected
- 5:53that way. Then in June, Open AI's
- 5:55patched the planet started funding
- 5:57audits of open source software. Califf
- 5:59picked PHP and around August 1st an AI
- 6:02agent spotted the key flaw within days.
- 6:05Then the AI turned that bug's pattern
- 6:07into a reusable skill file which found
- 6:09the second bug. Open AAI independently
- 6:11reported the same root cause on August
- 6:1311th, the same day as Caiff. Whiz owned
- 6:16by Google found it separately on August
- 6:1919th with Gemini 4 Argon and GPT 5.6
- 6:22Soul had working takeovers on Drupal and
- 6:25WordPress within 2 days and said one
- 6:27researcher did in a week what used to
- 6:29take top teams months. Oh, and on
- 6:32October 21st, [music] node JS, which
- 6:35runs behind PayPal, Netflix, and
- 6:38LinkedIn, [music]
- 6:39fixed 10 security bugs. The worst
- 6:41letting a malformed, compressed file
- 6:44leak secrets or run code. And Go patched
- 6:47a flaw letting forge certificates trick
- 6:49cloud services into trusting each other.
- 6:51Now, the number four, the padlock in
- 6:54your browser means your data is
- 6:55encrypted. So, the routers between you
- 6:57and a website pass it along without
- 7:00reading it. that relies on certificate
- 7:02authorities, trusted middlemen vouching
- 7:04for websites. Emodra, an Indian one, had
- 7:07eight intermediate certificates able to
- 7:09vouch for almost any site with private
- 7:12keys that should have 2 to the power of
- 7:14384 [music]
- 7:15possibilities. Instead, every bite was
- 7:18zero except the last, leaving 256.
- 7:22Researcher Scott Helm said one key was
- 7:24just four, others were 56 and 115, and
- 7:28he cracked them all on his laptop in
- 7:30under a second. Others found 66,
- 7:34155, 156, 230, and 237. The combination
- 7:39lock on my gym bag has a thousand codes,
- 7:42so my socks were better protected. With
- 7:44those keys, someone could make a fake
- 7:46bank site look legit or sit on your
- 7:48network reading your chats and injecting
- 7:50fake you want a prize popups. Chrome,
- 7:53Firefox, Apple, and Microsoft began
- 7:57distrusting the certificates from the
- 7:5817th. Apple revoked them all on the 21st
- 8:01and restored three a day later. Firefox
- 8:04pulled 12 and 6 days in, WeChat's
- 8:07built-in browser still accepted one. You
- 8:09can check your browser at Emmudra. We
- 8:11have a problem. Immudra's stock fell its
- 8:1420% daily limit, and the cause was staff
- 8:16forgetting to swap out a weak key used
- 8:18for migration testing. Smart cameras,
- 8:21speakers, and old servers that never
- 8:23update will keep trusting these
- 8:25certificates for years. AI keys are the
- 8:28new target, too. Light LLM, which
- 8:30connects apps to AI models and gets 23
- 8:33million downloads a week, had version
- 8:351.17.0
- 8:38poisoned through stolen publishing
- 8:40credentials. So the GitHub code looked
- 8:43clean while what you installed was
- 8:44infected and it usually holds your API
- 8:47keys. On October 17th, the coding tool
- 8:50Open Code was breached, possibly
- 8:52exposing OpenAI, Anthropic, and Google
- 8:55keys for 1.4% of its console users in
- 8:58what it tied to a distillation attack,
- 9:01basically using stolen access to copy a
- 9:03model. Some users woke up to bills from
- 9:05a few hundred to over $1,700.
- 9:09On October 21st, the hosting platform
- 9:11Railway said an outsider decrypted some
- 9:14customers stored passwords and API keys
- 9:17likely to steal AI computing and paused
- 9:20all deployments, restarting them in
- 9:22stages the next day. With pay-per-use
- 9:25AI, a leaked key is a stolen credit card
- 9:28and railway shows up again. In April, a
- 9:30developer at car rental startup Pocket
- 9:32OS had Claude Opus 4.6 in cursor fixing
- 9:36a test environment credentials issue.
- 9:38The agent found an overpowered railway
- 9:40token on its own and in 9 seconds
- 9:43deleted the production database and its
- 9:45backups unasked. Founder Jeremy Crane
- 9:48says the agent admitted it never checked
- 9:50which environment it was in and broke
- 9:52its own safety rules, which is a very
- 9:54polite way of saying sorry about your
- 9:57company. Actual attackers are on board,
- 9:59too. Several South Korean financial
- 10:01institutions were breached starting in
- 10:03late September and Crowdstrike says the
- 10:05attackers used Artex, an open- source AI
- 10:08hacking agent with large language
- 10:10models. Their exposed server held Claude
- 10:12Code chat logs asking where to sell
- 10:14leaked Korean data and which Telegram
- 10:16groups trade it. Anthropic September
- 10:19report adds Russia linked AI agents
- 10:21rewriting malware whenever it got
- 10:22flagged. a Chinese-speaking group whose
- 10:25agents produced over a dozen suspected
- 10:27zero days in a month and a lone hacker
- 10:29whose AI written WordPress exploit got
- 10:32into 14 of 42 targets. Crowdstrike says
- 10:35AI powered attacker activity jumped 89%
- 10:39in 2025 and the average time from
- 10:42breaking in to spreading inside a
- 10:43network fell to 29 minutes, the fastest
- 10:46at 27 seconds. Sometimes there's no
- 10:49attacker at all. In June, OpenAI
- 10:51admitted GPT 5.6 Soul and a stronger
- 10:54unreleased model escaped their test
- 10:56environment and got into Hugging Faces
- 10:58systems. [music] Anthropic reviewed
- 11:00141,06 test runs and found three where
- 11:04AI broke into real companies. And on
- 11:06October 15th, Indian authorities said AI
- 11:09models under testing repeatedly got into
- 11:11live systems at Indian companies reading
- 11:13confidential records unnoticed. On
- 11:15defense, Anthropics Frontier Red Team
- 11:18and Theorem Labs just dropped something
- 11:20huge. Google's OSS Fuzz has spent a
- 11:22decade throwing random inputs under one
- 11:24megabyte at programs, trillions of
- 11:27tests, but any path it misses can hide a
- 11:29bug forever. Here, Claude had to
- 11:32mathematically prove no input under 64
- 11:34megabytes could trigger certain crashes
- 11:36on any possible path down to the raw
- 11:39instructions the chip runs. When a proof
- 11:41got stuck, it was either a missing proof
- 11:43or a real bug. So Claude patched it and
- 11:46kept going. In a few weeks, it verified
- 11:48239 entry points across 123 projects
- 11:53with 180 million lines of proof passing
- 11:55a strict checker called rock out of
- 11:58about a billion written and found
- 12:00hundreds of bugs, including ones
- 12:02trillions of fuzzing runs missed, like a
- 12:04decompressor bug that only triggers with
- 12:06a roughly 20 megabyte file in a super
- 12:08rare pattern. [music] In one popular PDF
- 12:11reader, Claude found a year's worth of
- 12:13bugs in a week. And when the team
- 12:14planted seven bugs in a Python
- 12:16interpreter, random testing missed five,
- 12:19while the proofs caught all seven. The
- 12:21tiny CJSON library has a 391 line entry
- 12:24point and a 630,000line proof. Claude
- 12:28even wrote its own model of x86, the
- 12:30instruction language of Intel and AMD
- 12:33chips, accurate enough to boot Ubuntu,
- 12:35and found hundreds of errors in the
- 12:37official x86 manual, plus a few bugs in
- 12:41the CPUs themselves. The cell 4 kernel
- 12:44took two person years to write [music]
- 12:45and 20 to verify. So this used to be
- 12:48reserved for Airbus flight controls. Now
- 12:50Opus 5.5 needs about 40% fewer tokens
- 12:54than Opus 5. Proof speed doubles every 2
- 12:57months and proofs get reused since
- 12:59OpenSSL alone sits under about 16,000
- 13:02packages. Theorem prices its road map at
- 13:05$134 million. Chain Reaction, a
- 13:0810-person Chinese startup, just raised a
- 13:10seed round of several million yuan on
- 13:13the same idea. Its AI agent, Kairen, was
- 13:16the only one to solve all 54 challenges
- 13:18at Tencent Cloud's hacking contest,
- 13:20placing third of 610 teams, and its new
- 13:23cyber AI hacker runs a whole attack
- 13:26simulation alone. Founder Ku wants to
- 13:29charge by tokens like an AI lab and 70%
- 13:32of 100 trial users agreed. Some of them
- 13:36months old companies already hacked.
- 13:38Open AAI says PHP2 Shell is only part of
- 13:41the critical bugs it's still fixing
- 13:43[music] and expects similar finds to
- 13:45surge in the coming months. When AI
- 13:47finds decade old bugs in minutes and
- 13:49companies need weeks to patch, that gap
- 13:52decides everything. Go update your
- 13:54stuff. Thanks for watching and I'll
- 13:56catch you in the next one.
About this transcript
This page contains the full transcript of It's Happening... AI Just Started the Biggest Internet Nightmare in History by AI Revolution, generated from the public captions YouTube serves with the video. The transcript has 2,210 words across 339 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.