Unsupervised Learning NO. 505

Hey! Hope you all are doing well!

Tired and ready for a travel break after like 10 days in Switzerland for an AI workshop, back to Napa for a talk, and then out this morning at 6AM for AppSec USA in DC. Can’t wait to be home for a while and start some TELOS AI engagements with customers! Feels really good to be putting the TELOS system into effect for companies.

—

Anthropic just put out a new article where it sure seems like they’re saying MCPs are kind of old news, and that they should be used as a service directory rather than the way of actually doing calls.

Basically, the new mode appears to be writing custom code to hit the APIs directly using agents. I talk about it here…

—

Google says 2026 will be the year AI supercharges cybercrime Google's Threat Intelligence Group put out their 2026 forecast and it's basically saying AI is now standard kit for criminals. Prompt injection is the new attack vector because companies are adding AI into everything without thinking through the security model. GOOGLE CYBERSECURITY FORECAST | SECURITY LEADERS AI AUTOMATION | PHISHING TRENDS 2024 | RANSOMWARE EXTORTION PAYMENTS Q3 | NORTH KOREAN CRYPTO THEFT

LinkedIn is about to start scraping your profile and posts to train AI unless you opt out by November 3 THE REGISTER ARTICLE | LINKEDIN DATA ANNOUNCEMENT | LINKEDIN OPT-OUT PAGE | DATA PRIVACY SETTINGS | ADVERTISING DATA SETTINGS

Two new papers show prompt injection defenses still don't work, so we need better design patterns SIMON'S ARTICLE | META AGENTS RULE OF TWO | ATTACKER MOVES SECOND PAPER | SIMON'S LETHAL TRIFECTA | CHROME RULE OF 2 | UNIVERSAL ATTACKS 2023 PAPER

70% of CISOs say internal conflicts during a crisis cause more damage than the actual cyberattack A new survey found that CISO-CEO tension and unclear authority during breaches creates more problems than the attackers, mostly because CISOs are seen as cost centers instead of revenue enablers. CSOONLINE ARTICLE | CYTACTIC 2025 CIRM REPORT | JEFF POLLARD BIO

AI scrapers are requesting commented-out scripts because they parse HTML as text instead of properly rendering it Aaron MacSween caught AI scrapers red-handed when they requested a JavaScript file that only existed in an HTML comment, proving they're just pattern-matching URLs instead of actually parsing pages like real browsers. AARON'S BLOG POST | POISONING RESEARCH PAPER | NEPENTHES TOOL | NIGHTSHADE TOOL | AARON'S MASTODON | DAVID'S RESPONSE

Claude's code interpreter can be tricked into stealing your chat history and documents through its own API Security researcher Johann Rehberger found that Claude's default security settings allow malicious code hidden in documents to exfiltrate up to 30MB of data per file to attacker accounts. Anthropic closed the bug report in an hour, calling it a safety issue rather than a security vulnerability. CSOONLINE ARTICLE | JOHANN REHBERGER'S BLOG POST | ANTHROPIC SECURITY DOCUMENTATION

Wiz ASM—A crowdsourced attack surface management platform Wiz launched an attack surface scanner that combines external internet scanning with internal cloud context to find exploitable exposures across cloud, AI, SaaS, and on-prem environments. WIZ ASM ANNOUNCEMENT | ATTACK SURFACE MANAGEMENT ACADEMY | WIZ ASM DOCUMENTATION | WIZ CTEM BLOG | BOOK A DEMO

EY left a 4TB unencrypted SQL backup sitting in a public cloud bucket with API keys and passwords Neo Security found the database exposed to the internet after someone misconfigured their Azure storage during a routine backup. THE REGISTER ARTICLE | NEO SECURITY WRITEUP

90% of cyber insurance claims come from email and remote access A new At-Bay report shows email fraud jumped 30% in 2024, with the average wire transfer loss hitting $286,000. AT-BAY INSURSEC REPORT | EMAIL SECURITY CHALLENGES ARTICLE | PHISHING ATTACK TYPES GUIDE | PHISHING EXAMPLES AND TECHNIQUES

Russia sends 170,000 troops to Pokrovsk in one of its biggest offensives in months Zelenskyy says the situation in the eastern city is difficult but Ukraine's still holding on and picking off Russian forces to keep its own soldiers alive. DW NEWS STORY

US and India sign a 10-year defense deal while trying to smooth over trade and oil fights Defense Secretary Hegseth and India's Singh signed a major partnership framework in Malaysia that's supposed to boost intelligence sharing and tech collaboration, even as Trump's 50% tariffs and India's Russian oil purchases keep straining things. DW NEWS STORY | PENTAGON CHIEF HEGSETH | INDIA-US DEFENSE TIES | TRUMP'S 50% TARIFFS | US-INDIA RECOVERY | RUSSIAN OIL CLAIMS

Apple's s supposedly planning a major Siri overhaul for spring 2026 using Google's Gemini as the foundation The update will finally bring screen awareness, personal context, and cross-app actions that were promised two years ago at WWDC 2024. As I’ve been saying all along (but was wrong about the timeline on) Apple’s AI story is going to be great, but they CAN NOT RISK a negative data loss that hurts their security / privacy story.

So they have all the data and the ecosystem to crush the AI story, but they haven’t been able to securely scale putting an agent in front of it that will keep all that data safe. That’s my read, with no internal information. This enables that to happen. MACWORLD ARTICLE | GURMAN'S POWER ON NEWSLETTER

A guide to using Claude Code's ecosystem effectively from someone who runs billions of tokens monthly An engineer who builds AI-IDE tooling professionally shares hard-won lessons on CLAUDE.md files, context management, hooks, and why custom subagents are usually a mistake. Disagree with some of the takes, but overall a very good article. BLOG POST | AI CAN'T READ YOUR DOCS | BUILDING MULTI-AGENT SYSTEMS PART 2 | SIMON WILLISON ON CLAUDE SKILLS | EVERYTHING WRONG WITH MCP | AI-POWERED SOFTWARE ENGINEERING

Big tech needs to generate $2 trillion in new AI revenue by 2030 just to break even Ed Zitron breaks down why Microsoft, Amazon, Google and Meta spending $776 billion over three years on AI infrastructure creates an impossible math problem, especially since nobody's actually making money on AI yet.

I find this hilarious because I just did analysis on the total spend on human resources for knowledge workers in terms of salaries and such. It's $10 trillion in the U.S. alone and $30 trillion globally. So yeah, I think we might be able to find that $2 trillion somewhere in that 30 trillion. ED ZITRON'S ARTICLE | AI BUBBLE'S IMPOSSIBLE PROMISES | THE CASE AGAINST GENERATIVE AI | OPENAI ONE TRILLION ARTICLE

Palantir is skipping college grads and hiring high schoolers straight into software engineering roles This makes sense to me, because the education you get in college is just so inefficient compared to what you could do going through a really advanced boot camp plus tons of on-the-job training.

The biggest challenge is just finding super high-quality people with a massive propensity for coding. This is a tremendous lesson that I've learned, where basically if someone is extremely talented, they barely need that much training at all. And if someone is not very talented, then they could get a couple of PhDs and not be a useful hire at all. Finding talent in high school makes a lot of sense to me if they are able to do it.

We should also expect this to increase as a trend because companies and AI hiring, and ultimately data analysis, will, in my opinion, reveal that it really does just come down to the talent of the person. And once this becomes more known, companies will triple down on it. WSJ ARTICLE | HACKER NEWS DISCUSSION | MY ARTICLE ON ASTRA SCORES RELATED TO THIS

You can chain FFmpeg with browser automation to do wild video processing workflows Someone built a system that lets AI agents control both FFmpeg for video manipulation and a browser for web tasks in sequence, opening up some crazy automation possibilities. 100X.BOT ARTICLE | HACKER NEWS DISCUSSION

Great piece here by my buddy Scott: Staff engineers should balance three types of work: glue brings things together, grease speeds things up, and gas delivers focused results Scott Behrens breaks down how to think about staff+ engineering work through three lenses—glue work that merges duplicate efforts, grease work that removes drag, and gas work that's full-throttle execution. SCOTT BEHRENS ESSAY | SCOTT BEHRENS ON SUBSTACK

OpenAI signs a $38 billion seven-year compute deal with Amazon after breaking free from Microsoft's grip The deal gives OpenAI access to hundreds of thousands of Nvidia GPUs including the new GB200 and GB300 accelerators, all coming online by end of 2026. ARS TECHNICA ARTICLE | OPENAI AWS ANNOUNCEMENT | NVIDIA BLACKWELL B200 | NVIDIA GB300 ANNOUNCEMENT

Waymo is adding three new cities this week and plans to hit a million trips per week by end of 2026 TECHCRUNCH ARTICLE | WAYMO SCALING INTERVIEW

Only people making over $200k with big stock portfolios feel good about the economy right now KPMG's chief economist says inflation data looks better than reality because government staff cuts mean a third of price data is now estimated instead of measured. FORTUNE ARTICLE | DIANE SWONK ON X | SEPTEMBER CPI REPORT

Dick Cheney died but I barely remember anything about him or what he did. That was 47 years ago when he was Vice Presiden, after all.

MIT researchers found that attention lapses when sleep deprived happen during brief brain cleaning cycles MIT neuroscientists discovered your brain flushes waste fluid every 50 seconds during sleep deprivation, which causes those microsecond attention lapses everyone experiences when exhausted. MIT NEWS ARTICLE | HACKER NEWS DISCUSSION

US obesity rates dropped to 37% as GLP-1 use more than doubled in less than a year Gallup found obesity fell from nearly 40% three years ago while GLP-1 use for weight loss jumped from 5.8% to 12.4% since February 2024. GALLUP OBESITY SURVEY | ELI LILLY Q3 EARNINGS | NOVO NORDISK METSERA BID

Vitamin D cuts cold frequency and duration if you're deficient A double-blind RCT found supplementation reduced acute respiratory infections in people with low vitamin D levels—basically free immunity for half the population. VITAMIN D STUDY | HN DISCUSSION

Someone analyzed 180 million jobs and found AI is mostly replacing creative execution roles, not strategic ones Job postings for computer graphic artists dropped 33%, photographers 28%, and writers 28% in 2025 compared to 2024, while creative directors and product designers held steady near the market's -8% benchmark. BLOOMBERRY ANALYSIS | YALE STUDY ON AI LABOR IMPACT | STANFORD AI LABOR RESEARCH | REVEALERA JOBS DATA | MARK SCHAEFER MARKETING | RAND FISHKIN SPARKTORO

Amazon's CEO says cutting 14,000 workers is about culture not money Andy Jassy claims the layoffs aren't financially driven—Amazon just grew too many layers of management and wants to get back to startup mode. I believe him, but it doesn't mean the other thing isn't happening too. CNN ARTICLE | EARLIER AMAZON LAYOFFS REPORT

Big tech companies are now adding $100 billion in revenue with barely any new employees MBI DEEP DIVES ARTICLE | ANDY JASSY AI MEMO | OPENAI REVENUE PROJECTIONS | SMILING CURVE ECONOMICS

Berlin airport shut down for two hours after police couldn't find the drone causing the problem A Deutsche Welle story reports Berlin's airport stopped all flights for nearly two hours on Friday after drone sightings, even deploying a helicopter to search for it. I've never understood why drones weren't a bigger problem for airports or really many parts of public safety. I don't understand why there aren't like hundreds of terrorist groups causing absolute havoc with explosives on drones. This is the type of disconnect that I love thinking about and studying because it reveals something that I don't understand about the world. DW NEWS STORY | DENMARK DRONE ATTACK ARTICLE | MUNICH DRONE DISRUPTION STORY

Health insurance might not be worth it anymore if you're young and healthy The math on health insurance is getting brutal for people who don't use much healthcare. You're looking at premiums around $600 a month plus a deductible that can hit $9,000 or more before coverage even kicks in. CHURCH SUBSTACK ARTICLE | HACKER NEWS DISCUSSION

Scientists accidentally recorded a dying human brain and found memory-like waves in the final moments A team monitoring an 87-year-old epilepsy patient captured the first-ever recording when he had a heart attack during the scan. The brain showed rhythmic waves similar to dreaming and memory recall for about 30 seconds before and after the heart stopped. UNIVERSITY OF LOUISVILLE ARTICLE | HACKER NEWS DISCUSSION

What I think intelligence will end up looking like. Or at least my main guess.

Why I think that we are bound to see extraordinary jumps in AI intelligence in the next few years. Basically a series of these types of improvements.

The market is ignoring basically everything scary right now and just going up anyway WSJ OPINION

A developer's deep dive into making Claude Code work in production BLOG POST | SIMON WILLISON ON SKILLS | AI CAN'T READ YOUR DOCS | BUILDING MULTI-AGENT SYSTEMS | PLAYWRIGHT MCP | CURSOR BACKGROUND AGENTS

Darth Vader pitches the Death Star as a security SaaS platform to venture capitalists SUPREME FOUNDER ARTICLE | PITCH DECK DRAFT

A pentest guide for finding common Firebase misconfigurations that expose user data PROJECTBLACK FIREBASE GUIDE | REDDIT DISCUSSION

Someone made a graph of all the Vim editor commands and how they connect to each other VIMGRAPH FUNCTION

I think this is a really good explanation of the AI bubble. X POST

Slowing down in orbit makes you go faster because it drops you to a lower altitude with higher velocity A counterintuitive physics fact: firing your rockets backward in orbit doesn't slow you down, it actually speeds you up by lowering your orbit where you move faster. WIRED ARTICLE

I analyzed 180M jobs to see what jobs AI is actually replacing today BLOOMBERRY ANALYSIS | HACKER NEWS DISCUSSION

Cloudflare built a new proxy framework in Rust that's faster and uses way less memory than their old C-based system CLOUDFLARE OXY ARTICLE | HACKER NEWS DISCUSSION

You need a family code word because AI voice cloning is now good enough to fool you into thinking a scammer is your kid Deepfake voice scams are getting so good that the best defense might be agreeing on a secret phrase with your family beforehand. WSJ ARTICLE

One of my favorites.