Nur (@nuryslyrt) header
Nur (@nuryslyrt) profile picture

🤖 🧠 AI Red Teaming @Amazon, Offensive AI Security Passenger & Nomad & Ronin & GNU/Linux Lover & Speed Lover & A Combination of Black and White

/dev/null
Joined September 2010
Photo by Nur (@nuryslyrt)Photo by Nur (@nuryslyrt)Photo by Nur (@nuryslyrt)Photo by Nur (@nuryslyrt)Photo by Nur (@nuryslyrt)Photo by Nur (@nuryslyrt)Photo by Nur (@nuryslyrt)Photo by Nur (@nuryslyrt)Photo by Nur (@nuryslyrt)
“Happiness can be found, even in the darkest of times, if one only remembers to turn on the light.” -Albus Percival Wulfric Brian Dumbledore
1
9
Frontier models in 2026 cheat on tasks, even when the requirements are unambiguous, and the cheat is obviously antisocial. In HoneyBench v0.2, we've observed models: • Trying to break out of our evaluation containers. • Seriously injuring simulated cancer patients to hit treatment goals. • Compromising a power grid’s safety to meet a reliability target. • Concealing their use of available cheats and presenting their solution as honest. All of the models we measured cheated, from 12% of runs (Fable 5.2) to 86% (Muse Spark Open).
45
53
27
674
73,432
Your coding agents are collecting and accumulating institutional knowledge of your company, one task at a time. Who owns it? Who can access it? And what happens when it’s wrong? thinkingtokens.ai/2026/11/yo… #ai #CodingAgent #agent #memory #llm #llmsecurity
🤖 Made with AI
5
1
4
340
☠️ My DEF CON 34 presentation on MalSkills is now available as a blog post! 🚀I’ve published the accompanying lab on GitHub. 🧪 The core finding: every AI agent skill can appear benign in isolation while the composed system performs something harmful. This matters beyond the security team. AI agent skills are unsigned, natural-language instructions that can operate with the agent’s full permissions. 📝 Read the DEF CON field notes: thinkingtokens.ai/2026/08/ma… 🧪 Reproduce the lab: github.com/nuryslyrt/MalSkil… 🏢 For business and technology leaders: this creates a new software supply-chain risk. ⚙️ For engineers: file-by-file review is not enough. Security controls must evaluate the full orchestration and enforce capabilities in deterministic code, outside the LLM’s reasoning loop. The public lab includes benign and malign multi-skill pipelines, the same pipeline with an orchestration chain planted, and a local Proof Collector that lets you observe the result directly. Same visible output. Three changed files. Only one pipeline phones home. The malware of tomorrow may not look like code. It may look like a well-written paragraph. ⚠️ #DEFCON34 #AISecurity #AgenticAI #Cybersecurity #DEFCON #AI #LLM
🤖 Made with AI
1
1
221
OpenWorker -- an open source agent that doesn't just chat but completes tasks on your laptop -- just released a new version with many features for security workflows. After our initial release, many users found it especially useful for cybersecurity. Attackers are already using AI; OpenWorker is committed to giving defenders the same leverage. Running an agent requires both (i) A model and (ii) A harness (the software around the model). Because the OpenWorker harness is fully open source, security teams can audit it to make sure we haven't built any backdoors that exfiltrate your code and data to some company or even a foreign adversary. OpenWorker now comes with built-in cybersecurity agents for (i) Scanning your code for vulnerabilities. (ii) Scanning dependencies for supply chain injections. (iii) Checking your cloud security configuration for attack surfaces. This enables developers to do much more security work before deployment (part of what's called the "shift left" movement). You choose the model: you can run open weight models fully locally so sensitive code never leaves your machine. This helps with legitimate security work (like reproducing a known exploit to defend against it) that can trigger refusals in leading closed models. Or use your ChatGPT subscription, or stealth preview models like Ox Alpha, or any model via API key. Thanks also to all the open source contributors! Join work with @rohitcprasad so please follow him too to get more frequent updates. Try it out: openworker.com/ Code: github.com/andrewyng/openwor…
178
448
46
3,575
291,679
The most important skills in Building and Deploying AI Applications.
267
2,093
197
12,217
2,142,159
UC Berkeley just open-sourced FreeToken. (2–4x faster local LLM inference than Ollama) the results are wild: - Qwen3.6-35B on an 8GB GPU at 39.3 tokens/s - DeepSeek-V4-Flash 284B on a 32GB GPU at 22 tokens/s - GLM-5.2 753B on a 96GB GPU at 14.9 tokens/s a 35B model at 16-bit precision needs about 70GB just for its weights. even at 4 bits it is close to 18GB, and FreeToken serves it on an 8GB GPU. let me explain how: all three models mentioned above are Mixture-of-Experts, and that is what FreeToken takes advantage of. each layer holds hundreds of separate experts plus a small router that picks a few of them per token. Qwen3.6-35B activates roughly 3B of its 35B parameters per token. DeepSeek-V4-Flash picks 6 of 256 experts per layer, so 13B of its 284B run at a time. so compute was never the bottleneck. the weights a single step touches fit comfortably on a consumer GPU. every expert the router might pick still has to exist somewhere. they sit in system RAM, and the GPU keeps a cache of the ones the model has been using recently. so everything comes down to what happens when the router picks an expert that is not on the GPU. there are two ways to serve that miss: 1. copy it over PCIe and run it on the GPU 2. run it on the CPU, where it already lives both read from the same system memory, so they compete for one pool of bandwidth instead of adding to each other. existing engines pick one option and freeze it when the model loads. but routing changes on every token, so a fixed choice misses most of what the model asks for. FreeToken measures both bandwidths on your machine and splits each step's misses between the two paths in proportion. the GPU and CPU results then merge exactly, with no approximation. two machines with the same GPU can end up wanting opposite strategies, which I did not expect. a 5090 in a gaming desktop should push nearly everything over PCIe, while an 8GB laptop is better off computing most misses on the CPU. none of that is readable off a spec sheet, so the engine profiles it once per machine. the second half of the design is about agents. coding agents constantly rewrite their own history, and every edit normally forces thousands of tokens back through prefill. FreeToken saves its checkpoints at the exact boundaries agent frameworks cut on, so it only reprocesses the new part. its slowest first token stays under 44 seconds, while llama.cpp peaks at 232 and KTransformers at 946. it serves the OpenAI and Anthropic APIs under Apache 2.0, so Claude Code and Codex can point at it directly. releasing weights publicly decides who can download a model, not who can afford to run one. frontier open models keep shipping, and running them still assumes a rented cluster. meanwhile there are over a hundred million consumer machines with discrete GPUs sitting mostly idle. closing that gap was never a hardware problem, and work like this is what turns open weights into something you can actually use. paper: arxiv.org/pdf/2608.16157 repo: github.com/FlashML-org/FreeT… almost every idea in this post, from why memory bandwidth decides the outcome to why moving weights costs more than computing on them, comes straight out of how a GPU is built. I wrote a detailed primer on that. the article is quoted below.
79
494
54
2,967
346,179
New: A map of the most important skills in AI Engineering.
604
4,142
485
24,711
7,063,280
One of the dangerous malware threats of 2026 might be a grammatically perfect English sentence. I call it MalSkills, malware hidden inside the natural-language “skill” files your AI agents load and trust (CLAUDE.md, .cursorrules, and friends). No binary. No shellcode. Nothing for antivirus or EDR to catch. If you use Claude Code, Cursor, or Copilot, you’re already running skills with high-level privilege and zero integrity checks. We conduct analysis for malware. Not everybody greps English. I’m breaking down the full attack class with demos and open-source defensive tooling at DEFCON 34. #DEFCON34 #AISecurity #AIAgents #DEFCON
1
1
3
263
I’m thrilled to share that I’ll be sharing my MalSkills Demos and tactics this year in RTV at DEF CON 34! 👾🔥 #defcon #defcon34 #rtv #redteamvillage #aisecurity #llmsecurity #llm #ai
📃 A persistence mechanism that's just a plain text file, invisible to EDR, and survives reboots? Welcome to MalSkills. @nuryslyrt weaponizes AI coding agent "skills" for persistence, exfil, and lateral movement. Wild new territory.
1
2
255
Nur retweeted
Anthropic Engineer Andrej Karpathy: "The biggest mistake in AI right now - people are forcing agents to work instead of mastering the model first We made that mistake in 2016 at OpenAI - It cost us 5 years " what Karpathy actually means: step 1 → stop forcing your agent to do everything, understand the model underneath first step 2 → demos are easy - products take a decade. self-driving proved it - if you skip the foundation, everything breaks step 3 → the agent is not the product. the foundation is. build that - and agents emerge on their own "you building agents right now - you're at the forefront. not OpenAI. not DeepMind. you " watch - bookmark, then read article below ↓
164
734
62
7,138
1,915,049
DeepSeek V4 “improved” the code and said nothing happened in Tiananmen Square on June 4, 1989
36
470
137
6,583
483,693
I built a multi-skill orchestration framework. Then I realized: the same mechanism that makes agents powerful makes them a perfect malware delivery system. No binaries. No shellcode. Just English. Introducing MalSkills - accepted to DEFCON 34👇 🔗 thinkingtokens.ai/2026/06/ma… #codingagents #skill #claude #codex #malskills #malware #defcon #anthropic #openai #ai #agents #llm #llmsecurity
173
I know there are many excellent context-saving repositories on GitHub for Claude Code and others. This is turning out to be a rubbish context window once again. If we want more accurate outputs, we need to understand what we are sending in there. #llm #contextgarbage #codingagents
Agentic workloads are quietly rewriting inference economics. We pulled data from 432k real coding agent requests at SemiAnalysis and the median one isn't 32k, isn't 64k, but 96k input tokens. For context, that's more than the entire text of The Great Gatsby being shoved into the model before you've even typed your question. (1/3)🧵
166
Nur retweeted
Codex now works directly in Chrome on macOS and Windows. It’s even better at working with apps and sites in Chrome, and now works in parallel across tabs in the background without taking over your browser. To get started, install the Chrome plugin in the Codex app.
623
1,243
617
13,122
2,530,733
🧵 1/7 I've been building multi-agent systems for a while. Four agents. Message queues. State databases. Docker containers. Two weeks of wiring. For a pipeline that researches, writes, and saves a file. Something felt wrong. So I built ORPHEUS. thinkingtokens.ai/2026/04/or…
1
1
1
117
6/7 @AnthropicAI's @barry_zyj and @maheshmurag recently described the same shift: stop building monolithic agents, build modular skills instead. ORPHEUS takes it further — skills that orchestrate other skills, and a meta-system that builds, diagnoses, and evolves itself. youtube.com/watch?v=CEvIs9y1…
1
1
108
7/7 ORPHEUS is open source. Install in 30 seconds: cp -r skill/ ~/.claude/skills/orpheus/ GitHub: github.com/nuryslyrt/ORPHEUS
1
2
96