- Unsupervised Learning
- Posts
- UL NO. 416: Tracking AI Agent Activity, 400 SF Cameras, AI Sleeper Agents...
UL NO. 416: Tracking AI Agent Activity, 400 SF Cameras, AI Sleeper Agents...
Benign AI's "Many Eyes", OpenAI's Pentagon partnership, AI voice scams, Zuckerberg all-in on AGI, and more...

Unsupervised Learning is a Security, AI, and Meaning-focused newsletter and podcast that looks at how best to thrive as humans. It combines original ideas and analysis to bring you not just the newsâbut why it matters, and how to respond.
TOC
INTRO
How are you?
Tons of stuff going on this week as plans for the year start to solidify.
A number of paid talks are starting to fill in, with the earliest in February and the latest so far in October. Absolutely love the combination of getting the ideas out there, getting to travel domestically and internationally, and getting paid for it!
I put out the bundle of loot for attendees of my AUGMENTED AI course on Friday.
I uploaded a ton to my still unannounced project, and activity on it is already going strong. Cannot wait to fully launch this thing!
đĽI have a buddy looking for a Security SE position. Remote, US-based. Heâs a total and absolute gem. Heâs one of my mentors and the best SE Iâve ever seen in any field. He not only learns any product instantly, but he deeply understands the tech, the customer problem, and sales, so he is insanely gifted at connecting what the customer needs to the product or service in question. Send me a quick note if you or someone you know is looking for a superstar SE. He will get snatched up quickly. EMAIL HIM DIRECTLY
Ok, letâs goâŚ

MY WORK
Dark Visitors is a project that tracks AI agents doing various shenanigans on the internet and offering the ability to block them via robots.txt. HT to @securibee | by Dark Visitors | MORE
Super cool research on AI Sleeper Agents. Basically, agents that act cool normally but wait for a particular stimuli or moment to become vulnerable or take some other action. MORE | THE PAPER
đĄYou know how open source was supposed to provide âmany eyesâ and keep us safe? Well, benign AI agents will actually make that a reality. Auditing code. Crawling content for malware traps. Sending strange input to systems to try to trigger sleeper behavior, etc. The solution to malicious AI is, unfortunately, going to be benign AI tasked with finding it.
OpenAI is now partnering with the Pentagon for some projects, reversing its stance on military use of its AI. They are working on âa number of projects including cybersecurity capabilitiesâ (Bloomberg), but theyâre maintaining their âno-weapons developmentâ policy. MORE | MORE
Sponsor
2024 State of IT Operations Report
We surveyed 500 U.S.-based IT professionals to dig into where the biggest efficiency challenges are for ITOps teams in 2024. We learned how generative AI and workflow automation tools increase IT agility, reduce costs, and enhance teamsâ ability to simplify IT management. Download the report now to see whatâs working for high-agility ITOps teams, where low-agility teams are struggling, and how your team stacks up.
Scammers are now using AI to fake the voices of relatives in emergency scams, tricking people into acting fast without adequate time to scrutinize. The FBI has logged over 195 complaints about these "grandparent scams," with victims losing nearly $1.9 million from January to September 2023. Tell your loved ones about these! MORE
China has been getting Nvidia chips despite a US ban that was meant to stop that from happening. Theyâve been going through smaller suppliers, circumventing restrictions imposed in 2022 and 2023. MORE
San Francisco is going heavy on surveillance, evidently. They just installed 400 license plate readers across the city as part of the Flock Safety camera system. The police chief says it will help track down criminals, citing that 70% of crimes involve vehicles. Iâm honestly for this kind of thing, despite the fact that itâll have downsides. I just want there to be proper use and oversight. MORE
Advisories
đ¨Ivanti Directive Issued â U.S. federal agencies have been ordered to patch a critical Ivanti software vulnerability. | CRITICAL | MORE
â ď¸ Cybercriminals are exploiting TeamViewer to launch ransomware attacks by leveraging leaked LockBit builder tools. Huntress Labs' analysis of compromised endpoints revealed that attackers gained access through TeamViewer, attempting to deploy ransomware via a DOS batch file. MORE
Incidents
â ď¸ UK Councils Cyberattack â Three UK councils are grappling with a cyber incident that's knocked public systems offline. | SEVERITY: HIGH | RESPONSE: Systems isolated, no customer data breach found yet. MORE
đ¨ Chinese Espionage Campaign â Chinese hackers have been exploiting a VMware vulnerability for two years undetected. | CRITICAL | CVE-2023-34048 MORE
Vulnerabilities
𪳠GitHub Key Rotation â GitHub just rotated critical keys due to a high-severity vulnerability that exposed credentials. | HIGH | CVE-2024-0200 | CVSS Score: 7.2 MORE
𪳠Critical Vulnerabilities Patched â VMware and Atlassian have released patches for newly disclosed critical vulnerabilities. | CRITICAL | CVE-2023-22527, CVE-2023-34063 | CVSS Scores: 10, 9.9 MORE
TECHNOLOGY
A recent study by Boston Consulting Group shows that consultants using ChatGPT-4 significantly outperformed their AI-less peers in various tasks. The experiment involved 758 consultants and found that those using AI completed 12.2% more tasks, did so 25.1% faster, and produced 40% higher quality results. MORE
đĄThese are beginner numbers because this is all just starting. I think the big change in hiring in techâand companies in generalâstarting in the next 2-5 years will be letting go of the bottom 75% of performers (or just not rehiring them after attrition), and competing for the top 25%.
And within that group, the competition will be fierce for the top 1-10% who are gods with AI. Why? Because when theyâre augmented by even just Copilot or ChatGPT theyâll be superhuman. But as agent frameworks start to take over, they wonât be a Human + AI pair. Theyâll be a Human + AI Team pair. And that AI team might be hundreds or thousands of people behind a single person.
In other words, the competition for jobs, starting in the next few years will be against a top 10% performer whoâs backed by a farm of AI Agents, which gives them the output of 10-1000X that of a non-augmented, normal employee. Itâs no competition. And this is who companies will still be hiring. Everyone else, moving along the scale of competence over time, will be increasingly unemployable.
Thousands of AI Authors on the Future of AI. Super cool project that surveyed thousands of published authors on what they thought was coming in AI, and when. I think they were far too conservative, which I think is due to their academic bias. In other words, they seem too safe and sane to creatively imagine how fast this stuff could actually move. Which is also why so many academics were blindsided by November 2022. Still, I think the paper set up the questions pretty well, and itâs still interesting to see that many opinions in one place. MORE
đĄThe hardcore academic âMLâ types are the people Iâve seen be the most wrong about AI and where itâs going. At least in my opinion; juryâs still out of course. The problem is the disconnect between the culture of academia and the insanity that is GenAI. Academics are high in rigor and caution, which is awesome for some things, but itâs a hindrance if youâre trying to think big and crazy. And big and crazy is whatâs needed to play in the current game.
My recommendation is to think carefully about where you are, and where the people you follow are, on the scale of CreativityââRigor.
Donât listen much to people who are like âAGI is 10+ years away, if ever.â Or, âWhat we have isnât even real AI.â Or, âYou canât trust AI because it literally just makes stuff up.â People saying such things in an absolute sort of way are likely to either be low in OCEAN Openness and/or an academic.
Donât bring math to a poetry contest, and donât bring pessimism to an art contest.
Mark Zuckerberg has pivoted again. He was all about metaverse, and then he kind of went the AR way with Lex on his podcast, and now heâs all in on open-sourced AGI. Heâs doing a massive acquisition of Nvidia's H100 GPUs, expecting to own over 340,000 by year's end. 2024 is going to be ridiculous. MORE
I agree whoever said we shouldnât call a model open source if they only release weights.
Itâs not fully open unless you have the data and methodology as well.
Still, very cool.
â á´ á´É´ÉŞá´Ę á´ÉŞá´ssĘá´Ę âď¸ (@DanielMiessler)
6:58 PM ⢠Jan 18, 2024
Google's Circle to Search simplifies finding info on your phone by letting you circle an item on-screen to instantly search for it. Canât wait for Apple and everyone else to copy this. Super cool. MORE
Shining black light in public places (FAR-UV) could help prevent the spread of airborne diseases, potentially reducing the likelihood and impact of pandemics. Studies show that far-UV light can kill 99.9% of coronaviruses and other pathogens in the air, offering a passive defense against a wide range of respiratory viruses. MORE
Wenquai slashed their AI costs dramatically by optimizing Mixtral with GPT-4. They managed to reduce their daily AI expenses from a steep $100 to less than a dollar. MORE
Apple finally passed Samsung in global smartphone sales last year. Despite a general market decline, Apple's shipments rose by 3.7 percent, while Samsung's dropped significantly by over 13 percent, contributing to Apple's lead. This is my surprised face. MORE
HUMANS
The FDA just cleared DermaSensor, the first AI device that can detect all major skin cancers, aiming to improve early diagnosis. The device, which uses elastic scattering spectroscopy to analyze skin lesions, showed a 96% true positive rate for detecting skin cancers in a clinical trial. MORE
The self-checkout trend is hitting a wall as stores like Walmart and Target scale back or ditch the machines after facing increased theft and higher labor costs. Dollar General's CEO admitted they've over-relied on the tech, and now plan to boost staff numbers at checkouts. MORE
đĄI find it fascinating how sometimes tech and various movements try to jump too far ahead, too quickly, and then get pulled back. Sometimes only for a second, and sometimes for a long time. Work from home. Self-checkout. AI?
South Korea just rolled out a new visa for digital nomads, aiming to attract remote workers and boost its economy. The visa allows foreign residents to stay for up to two years, provided they earn over $65,860 annually and have comprehensive health insurance. MORE
Germany is doing something similar. They made it easier to get citizenship, aiming to attract global talent to fill job shortages. The new law reduces the residency requirement for naturalization from eight to five years and opens dual citizenship to all, not just EU and Swiss nationals. MORE
IDEAS & ANALYSIS
Everyone Should be a Thinker
One of my core beliefs is that itâs a stain on humanity that only certain people are considered to have thoughts worth sharing. I know this is just because weâre a young species, and it takes time to advance as creatures, and as a civilization. I get that. But itâs weird how civilizational retardation like this is considered normal while youâre living in it.
If you ask the average person what they think about the most important things in the world, like how to self-govern, free will, moral responsibility, the nature of reality, etc., theyâll blush and smile and make some sort of self-deprecating gesture. âThatâs for the smart people to think about.â They think itâs for special people, like public intellectuals, people who write books or go on TV. Itâs a travesty.
Human civilization will reach some modicum of advancement when itâs expected that every human on the planet is educated and empowered enough to not only have their own thoughts, but to believe those thoughts to be worthy of sharing. The fact that the percentage of people who believe that today is likely some obscene fraction of 1% should embarrass us all.
NOTES
Iâll be camping for the Apple Vision Pro at the Burlingame store on February 1st. I normally camp in the summer, so February will be a different experience for sure. If youâre insane like me, come say whatâs up.
Loving this Classical album, Pamart: PLANET GOLD, and Iâm sadly not much of a Classical person. MORE
DISCOVERY
đĽMoving from a Knowledge Economy to an Allocation Economy. MORE
You wonât be judged on how much you know, but instead on how well you can allocate and manage the resources to get work done.
đ ď¸Galah: an LLM-powered web honeypot using the OpenAI API | by Adel Karimi | MORE
đ Pfuzz â A Unix-style web fuzzer for finding security vulnerabilities. | MORE
đ ď¸ LAST - Scans code for security issues using OpenAI from the command line. | by Latio Tech | MORE
đ aifs â An AI filesystem tool for easy local semantic search. | by KillianLucas | MORE
Culture Change at Google (The Employee Isnât First Anymore) MORE
Navigating American healthcare might not require insurance, as paying cash can sometimes be cheaper and more flexible. MORE
How People Left Twitter, and How Itâs Going MORE
Powerful DALLE-3 Art Prompts MORE
đ ď¸ TweetFeed's return for the latest Indicators of Compromise shared by the infosec community by Daniel LĂłpez | MORE
A Search for More ChatGPT/GOT-3.5/GPT-4 âUnspeakableâ Glitch Tokens by MORE
Top Hacker News Books of 2023 MORE
RECOMMENDATION OF THE WEEK
Hereâs a cool heuristic for gauging your own happiness.
Pay close attention to how the success of your friends makes you feel. Not intellectually, but viscerally. Immediately. Within 1 second of seeing evidence that theyâre crushing it.
If it makes you smile uncontrollably, fist pump, and want to text them and hype them up, that means youâre healthy. Congrats.
If it stings, gives you a sinking feeling, or makes you angryâŚI recommend you talk to a therapist. This will destroy not just your relationships, but your life in general. Itâs actual poison.
APHORISM OF THE WEEK
The language of friendship is not words, but meanings.
â
Thank you for reading.
UL is a personal and strange combination of security, tech, AI, and lots of deeply human content. And because itâs so diverse, itâs harder for it to go as viral as something more niche.
So if you know someone weird like us, please share it with them. đŤś
Yours,
