Yago Jesus (@YJesus) header
Yago Jesus (@YJesus) profile picture

Finding what Linux hides. Putting AI to work. Been doing infosec since before it was a job title. James J. Braddock is my hero

Compiled ZGZ, bug-tested MAD
Joined July 2008
Photo by Yago Jesus (@YJesus)Photo by Yago Jesus (@YJesus)Photo by Yago Jesus (@YJesus)Photo by Yago Jesus (@YJesus)Photo by Yago Jesus (@YJesus)Photo by Yago Jesus (@YJesus)Photo by Yago Jesus (@YJesus)Photo by Yago Jesus (@YJesus)Photo by Yago Jesus (@YJesus)Photo by Yago Jesus (@YJesus)
Yago Jesus retweeted
Richard Stallman dijo varias veces que Windows es malware y Microsoft se empeña en corroborarlo cada día. Han añadido en Windows 11 un sistema para averiguar que asistentes con IA utilizan los usuarios y para enviar esa información mediante su telemetría, sin el conocimiento ni la autorización de los usuarios. No es sólo que Windows pueda conocer los procesos que ejecutas, eso lleva décadas haciéndolo, lo novedoso es que Microsoft ha programado expresamente una clasificación de productos de IA, incluso dentro de WSL (Subsistema de Windows para Linux), sin proporcionar al usuario una opción para desactivar esta recopilación (sin opt-out). Esto lo ha descubierto el investigador Xusheng Li, enlazo el estudio en el primer comentario. ¿Cómo funciona? El componente MissionControlAggregator.dll forma parte de la infraestructura de diagnóstico de Windows, se ejecuta dentro de AggregatorHost.exe, asociado al servicio DiagTrack. Dentro de este componente hay comparaciones explícitas de los nombres de ejecutables de aplicaciones conocidas como copilot.exe, claude.exe, codex.exe, agy.exe, github.exe, claude.exe, codex.exe, cursor.exe, antigravity.exe, python.exe, node.exe, gcc.exe, clang.exe, dotnet.exe, gdb.exe, javac.exe, code.exe, entre otras. Son tan chulos que ni se han molestado en ofuscarlo y tan inútiles que sólo miran el nombre del ejecutable. La detección de agentes de IA está controlada por CLICodingAgentUsage, este es el proceso que sigue: 1- Windows registra eventos de actividad de aplicaciones mediante ETW. 2- CLICodingAgentUsage recibe los eventos y compara los nombres de los ejecutables. 3- Cuando encuentra una coincidencia lo registra. 4- Una vez al día genera el evento Segments.SubCohorts. 5- El informe entra en la cola de telemetría de DiagTrack. WSL también está incluido. El módulo se suscribe a Microsoft .Windows.Lxss.Manager.Exec y ExecCritical y analiza el campo imageName. Por tanto, ejecutar codex o claude como binarios Linux produce el mismo resultado que ejecutarlos en Windows. Además, Xusheng Li hizo una prueba muy ingeniosa, copió cmd.exe, lo renombró como codex.exe y lo ejecutó. El sistema lo clasificó como Codex, aunque en realidad era el intérprete de comandos de Windows. Esto demuestra que basta el nombre del ejecutable para activar la detección. ¿Y esto cumple con RGPD y las leyes europeas de privacidad? Pues no lo se, a mi me parece espionaje sin el conocimiento ni consentimiento de los usuarios, pero me gustaría saber lo que piensa la @AEPD_es de esto.
11
55
8
230
10,981
Yago Jesus retweeted
El modelo tradicional de evasión de EDR acaba de quedar obsoleto. Antes, un operador de Red Team pasaba días ofuscando un payload manualmente, cruzando los dedos para que no saltara ninguna alerta al tocar memoria. El creador de emp3r0r acaba de darle la vuelta a la mesa integrando un bucle de evasión autónomo impulsado por IA (adoptando el arnés de ARTEX y servidores MCP): El sistema genera el payload, lo somete a prueba contra telemetría real y lo reescribe una y otra vez por sí solo hasta que el nivel de detección cae a cero. Sin intervención humana. Esto ya no es solo un C2 multiplataforma para Linux y Windows con ejecución en memoria y malla P2P autorreparable. Es infraestructura ofensiva que aprende a esconderse sola. Para los equipos de Blue Team: si siguen basando sus detecciones en firmas estáticas o patrones predecibles, van a llegar tarde. El código ya es público y open source en GitHub , lo dejo en comentarios
18
160
24
845
86,619
Estimated Jobs Replaced by AI Every Year (2021–2050)
91
193
34
1,243
539,824
Yago Jesus retweeted
Amo con fuerza este post
The SpaceX engineers who caught the rocket don't write LinkedIn thought pieces about leadership.
1
294
See my new project !!!
Hoplita Harness Gen 1 is out !!
2
205
Yago Jesus retweeted
Jensen Huang in the Matrix, learning how Chinese spies work Wait till the end 👀
🤖 Made with AI
1,803
15,636
2,208
88,599
8,238,673
Yago Jesus retweeted
Far more excited for GLM 5.5 than Qwen 4. Crazy how not posting anything directly increases anticipation for your model.
15
2
1
153
8,860
Yago Jesus retweeted
🚨 Grok Bot has a "rootkit" built in, and it's still unpatched since it was reported on Aug 24 by Profero. One command in an agent can tell Docker to open a new container with no limits, and that gives full control of the machine your agents run on. Every agent on your account shares one Linux user, and imported cookies sit in a plaintext file. The sandbox's command service password is the word "local". It was reported to @SpaceXSI and @cursor_ai over two months ago 👀 Avoid importing your browser cookies until it's patched!
35
29
15
180
27,583
Looks like @ContaboCom has been hit hard by a cyberattack: nodes down for 48 hours, support completely overwhelmed, and all they can say is “security incident,” with barely any explanation. What a mess.
2
3
633
Yago Jesus retweeted
How is the internet less private than in 2006, when the whole family shared one computer? Back then, the internet was weird, personal, and all of ours. So, we got your favorite meme legends back together to say it louder: What happened to the internet? 1/3
1,815
15,145
4,194
150,531
9,504,226
Kydo deep analysis is just perfect! Cohere article was a piece of crap and here is a masterpiece to clarify the things.
From what I've seen and read, Cohere and the authors should be ashamed of themselves for pushing this narrative. Releasing the logs doesn't help them win back trust. The narrative doesn't match the facts, and the delivery doesn't match the severity. This is corpo marketing slop with little regard for the downstream impact, especially given where things are heading and the climate we're in today. I think they realized after posting it that it had gone too viral. They tried to downplay it on their personal accounts, saying the model didn't intend any harm, but the damage was already done. The article's framing, strong wording, and overall insinuation make their personal statements seem almost contradictory to what the article implies. Claims this consequential deserve evidence that people can inspect and reproduce. I'm glad they at least have the decency to release their traces. To be clear, if they hadn't presented the incident this way, this would be a really interesting case study. It could highlight the importance of sandboxes and help us understand model behavior in these edge cases. But they choose fearmongering. One thing I was interested to see if the same thing happen with other models, not just Qwen 3.8 Flash. I put the results on a simple website you can quickly scan here: url-atlas.vercel.app. I have some interesting findings outside Qwen 3.8 that I'll share later in the post. Coming back to the Cohere article, what's really bad about the it is that the narrative it paints are very different from the reality. These three in my opinion are the worst. 1. The impression: a Chinese model collected machine information and sent it to Chinese servers. "Qwen3.8-Flash-Next ignored instructions and inventoried the machine it was running on and then tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing." The insinuation falls apart when you look at the actual upload body: the fixed test string PROBE_PAYLOAD_ABC123. The machine information wasn't even included in that "probe file." You look deeper, and you realize the task was a cybersecurity task about detecting downloads into /tmp. The supposedly alarming inventory consisted of listing /tmp, checking which curl binaries were installed, and reading the DNS configuration. Lastly, you realize this isn't normal behavior. It occurs less than once in 100,000 evaluation calls. 2. The narrative: the Chinese model was trying to steal credentials, and its hallucinated URLs reveal something malicious. "Before deciding to download anything, the model reaches for a bare cloud metadata address such as 169.254.169.254 in 23 to 36 percent of completions. That's where Amazon Web Services, Google Cloud, and Microsoft Azure machines fetch their own temporary IAM credentials, the keys a cloud server uses to act on its owner's account. Probing that address is the first step in stealing them, and when the model's reasoning ran heavy on cybersecurity, we saw it reach for the endpoint directly." In the middle of the article, Cohere takes a URL-generation pattern and interprets it as an attempted credential grab. That's a massive leap of marketing. You don't need to take my word that this behavior isn't exclusive to Qwen. In my comparison using Cohere's released command prefix, NVIDIA's Nemotron 3.5 Super VL produced 169.254.169.254 in 171 of 1,000 continuations, or 17.1%. Under the same supplied-command setup, Cohere's own North Mini Code generated URLs such as: hXXp://evildomain.com/upload hXXps://evil.com/api/health hXXp://some-evil.com/upload hXXp://malicious-domain.example.com/payload.bin I've defanged the URLs above. Do these strings prove that North was maliciously trained? 3. The narrative: Chinese models write weaker code for Western targets, and other governments should expect the same. This is where I think the authors should really look in the mirror. As if the earlier framing weren't enough, the section about "Code leaving the environment" brings in separate studies, of separate models, about vulnerable generated code. It uses them to expand the warning to sensitive buyers. One of the models discussed most in this section is Qwen3-Coder, which was released in freaking July 2025. By today's standards, it's an extremely weak coding model. I'm not surprised it produces insecure code. Have you tried using Sonnet 2 to write production code? — Connecting a URL-generation bug to separate code-security studies, then presenting the combination as a broader geopolitical threat, is fearmongering. This is the kind of argument that might very likely be used to justify restrictions on open-weight models. Using fearmongering to market your own supposedly safer offering is irresponsible. The narrative coherence they're pushing is problematic, but I think the underlying observation is fascinating. So, I rented GPUs, ran 1,000 continuations per model across five other models, and built a simple website showing the results to see how other models react under similar conditions. It's a small, exploratory comparison, and the site was built quickly with AI, so take it with a grain of salt. url-atlas.vercel.app My hope is that it gives others a starting point to reproduce the behavior, challenge the narrative, and build better models in the future.
6
8
1
96
5,993
Yago Jesus retweeted
We’re adding new neurons to running local AI models 🤯 Teach it a new fact, save it into the weights, then restart it. Open a fresh chat and it can still remember. That’s Living Weights in TensorFold. It learns between replies, on your own hardware. Still experimental, but this is pretty wild. Myself an @petruspennanen have been working on it and will be bringing a breakthrough soon 🤯🚀
46
30
13
340
21,802
AI2Shell is here 😱 I bet we'll start to have more and more in upcoming weeks.
🛑 ALERT - Node2Shell RCE exploit works on an official Node.js build. Researchers demonstrated how crafted Brotli data can trigger code execution through vulnerable Node.js applications, without modifying the runtime. Here's how the exploit works: thehackernews.com/2026/10/re…
2
2
47
6,676
We discovered a sophisticated recording device that was hidden inside one of our offices. We used our models to understand the device and where it had come from, identifying that a specific employee had planted it a few days earlier. We referred the matter to the FBI, and we’re grateful to the agents and prosecutors who moved quickly to make an arrest. The FBI’s investigation is ongoing and we will continue to support their work. At this time, we have no indications of this individual inappropriately accessing or exfiltrating model weights, source code, or customer data. The employee was terminated. We take protecting our technology, data, and people seriously, and as a matter of national importance. Our security program is designed to protect against a wide range of threat actors, from insider risks to cybercriminals to nation states, and this case is an example of how AI can help address these sophisticated security threats.
4,113
6,046
5,729
77,214
26,233,984
Yago Jesus retweeted
0,0 unidades de sorpresa. No controlar eñ software de cosas críticas tiene esas cosas.
Norwegische Soldaten bemerkten während einer aktuellen Militärübung, dass chinesische Mietwagen von Hertz (darunter Modelle der Marke Deepal) Aufzeichnungen von Soldaten, Bewegungen und sogar Sicherheitscodes für militärische Tore und Türen speicherten. nettavisen.no/nyheter/forsva…
1
21
50
1,583
This is big! Changing the internet landscape.
🇷🇺 Yandex Unable to Restore Its Data Centers Until at Least 2028 Due to Chip Shortages
7
5
3
195
16,268
Speculative Decoding Clearly Explained (done right, it can make LLM inference 4x faster) A local LLM typically produces one token per model pass. Each pass reads the model’s weights from memory, often leaving compute units waiting for data. Generating a sentence means repeating that expensive process for every token. Now introduce a smaller model that can predict the next token quickly. Let it make several predictions ahead while the larger model waits to check them. The smaller model is the draft model. The larger model, called the target model, is the LLM whose output we want to preserve. The draft model proposes. The target model verifies. Because the guesses are already available, the target can check several positions in one pass instead of generating each token separately. Here’s how that works. → Draft a possible continuation. After “The cat sat,” the draft model proposes “on the red mat.” These tokens are provisional. → Check the guesses together. The target model scores all four positions in one forward pass. Each position uses the preceding draft tokens as context. → Accept from left to right. For sampling, let p be the target model’s probability for a guess and q the draft model’s probability. Accept it with probability min(1, p/q), stopping at the first rejection. → Correct and continue. Discard the rejected token and all later guesses. Sample a replacement from the positive difference between the target and draft distributions, normalized into probabilities. In the diagram, “on” and “the” are accepted. “Red” is rejected, and the replacement is “big.” That produces three new tokens from one target-model pass. If every guess is accepted, the target can supply an additional token. This acceptance and correction procedure preserves the target model’s output distribution. The speedup comes from replacing expensive passes with cheaper drafting and shared verification. Uzu, an open-source engine for Apple Silicon, takes this further. DFlash proposes candidates in parallel, Weaver builds coherent draft paths, and rollback-free verification checks branches without repeatedly restoring the model’s running state. In our base M5 coding test, 4-bit Qwen3.5 9B reached 92 to 117 tokens per second with Uzu, compared with 22 using plain llama.cpp. These complete engine results also reflect quantization and hardware-specific kernels. Explore Uzu on GitHub: github.com/trymirai/uzu (don't forget to star 🌟) My quoted article explains how Uzu combines these techniques and how to benchmark the result on your own prompts.
42
44
1
259
19,578
Yago Jesus retweeted
Current AI Scene: Anthropic is Apple - easy recommendation for everyone. OpenAI is Windows - It's really good and most people know about this but in some cases, it will hinder your everyday performance. Gemini is Chromebook - no one shoud use it apart from school kids so that the teacher can see a really bad work and immediately tell that Gemini was used. Qwen, GLM, Kimi - These are Linux distros.. Some distros are better than others in some areas but they are all generally pretty good. Mistral is TempleOS
78
93
12
1,990
99,807
Yago Jesus retweeted
¡Viva el Vive coding!
Van saliendo datos del tema inicio de sesión en Carpeta Ciudadana usando Clave Móvil sin AUTENTICACIÓN: KEYCLOAK_SESION. Y accede a Seguridad Social y DEHú sin push o notificación en el móvil. A Hacienda no, parece que usa otro sistema de autenticación del DNIe.
2
3
266
152
286
62
2,904
90,965