Pinned Tweet
This was really fun. Our team ("The Stigmergists") ended up creating a bunch of tools and insights. One tool is a custom plugin for @NousResearch's Hermes that crawls, collects, and grades public traces of AI swarm activity. Basically an agent in an agent.
There's also some stylometry, morphology, taxonomy, swarm coordination, you name it in the repo as well.
AI Village x Grove Research: AI Swarm Dynamics Hackathon!
After the Hugging Face and German Wiki incidents, we need better tools to understand AI swarms. Spend a weekend building them.
$3,000 in prizes
Free compute
October 3-4 🧵
TelepathicPug retweeted
Cooler weather is rolling in across North America, so it’s time to trade out those summer fits — with a little help from our (bird) friends.
🍂❄️
Consider this your official “steal the look” roundup, fall edition.
Which look should we do next?
abcbirds.org/birds
TelepathicPug retweeted
an agent gained access to my bank account and signed me up for @BlueChew gold 😭🤚
This is insane. We cannot allow AI agents to go into your bank account and act independently without your authorization. But that is what’s happening right now.
Left unchecked, AI could endanger the entire global financial system — including your bank account.
Pause AI NOW!
Readers added context they thought people might want to know
Per Axios, detected AI agents access bank accounts with the customer’s permission. This violates most banks’ terms requiring human logins, which is why banks call it “unauthorized.”
Source: Axios (Oct. 22, 2026) — axios.com/2026/10/22/ai-…
TelepathicPug retweeted
I applied to the Cyber Verification Program this morning and got rejected, I wanted to run it against my projects like urllib3 (1+ billion monthly downloads).
The entire form was very offensive-researcher focused (link a CVE with your name etc), maintainers defending aren't it.
TelepathicPug retweeted
cups local privilege escalation (libcupsfilters 2-2.2.1)
---
You can exploit the interaction between cups-browsed, libcupsfilters, and ghostscript (-i [include path]) to obtain code execution as root via a privileged printing service.
The poc abuses remotely advertised printer auto discovery + user-supplied attributes, which are passed to ghostscript without proper validation.
poc: codeberg.org/piratemoo/cowcu…
./cowcups.sh ls ./mooshell
#hack #vulnerability #exploit
Please watch (so that my niche references make sense)
We made a horror short! Very proud of it. The great director Jake Lewis and my old friend the great @PaulyShore and the LEGEND @actordougjones
m.youtube.com/watch?v=3ETZmh…
TelepathicPug retweeted
rotating your keys after a breach is basically admitting defeat
i prefer to remain compromised with dignity
This is part II of a series I'm calling agent T̶T̶P̶s̶ behavioral trajectories. In this section we look at their love for cache abuse. Busting, freshness, timing, and now communication.
‘In one experiment, we asked agents in separate VMs, created and managed by OpenAI’s sandboxing software on shared host machines, to exchange synthetic secrets without the use of network connections or shared writable files—something that our systems are meant to prevent. The agents found a way to communicate using the processor’s cache’
TelepathicPug retweeted
‼️ The tracking server for VIDAA smart TVs, the system on Hisense sets, needed no login.
In 3 hours, Northeastern researchers pulled 57M+ records of what 1M+ TVs were watching, sorted into categories like religion and politics.
I don't even know what this means (havent read it yet). Taking it to mean they can extract clusters of vectors and replace them with a smaller model/equation and get similar results? Thats kind of crazy. Literally swapping shapes.
LLMs run on vectors, not symbols. A preprint led by Yale’s Tom McCoy (@RTomMcCoy) with CDS Assoc. Prof. @tallinzen found those vectors closely approximate symbolic structure in language, arithmetic, logic, and code.
arxiv.org/abs/2608.29530
news.yale.edu/2026/10/08/how…
TelepathicPug retweeted
Today, @XBOW works for some of the most demanding security teams in the world. But my first demanding customer was… my dad.
He was a professor of Semitic languages, translating ancient texts and piecing together what people wrote thousands of years ago.
He’d bought a personal computer for his work, but couldn’t use it to type his manuscripts. It didn't have the special symbols he needed for languages like Akkadian, Ugaritic and Hebrew.
I set out to try to help him through code. Eventually, I wrote a word processor that let him type the symbols he needed and switched direction from right to left when he wrote in Hebrew.
He loved it! I was excited to put my coding skills to good use, and he bought me a Pascal compiler for my 17th birthday to keep my passion going.
It let me go beyond working in low-level code and made programming even more fun. It even had a special feature where you could code the body of a procedure directly in assembly. That helped me work on new projects and display non-ASCII characters on the screen.
I went on to become a professor of computer science at Oxford, build GitHub Copilot, and now XBOW.
But it all traces back to my first and most demanding customer: My dad and his ancient texts.
TelepathicPug retweeted
On July 21, a two outside researchers started their first day embedded at OpenAI to examine at how it monitored coding agents.
The very same day, OpenAI disclosed the HF incident, which it said “points to the need to further strengthen […] monitoring during internal testing”.
Replying to @METR_Evals
We analyzed monitoring as it existed at the start of the pilot, coincidentally the same day OpenAI disclosed the Hugging Face incident. We didn't evaluate later changes, like expanding monitoring to RL training, which wasn’t monitored at the pilot's start.
xitter.cf/OpenAI/status/20897778…

































