← Feed

Posts

• View post

  Nice post-mortem on 1600 agent-derived vulnerabilities. Takeaways: * benchmarks are saturated * multi-step attacks are commonplace * threat is rising * offense drives defense https://xbow.com/blog/we-ran-1060-autonomous-attacks

• View post

  SharedRoot; Escaping the Claude Cowork sandbox Security is, at its core, a design problem. There will always be vulns and misconfigurations; design is what protects us all. The whole "Johnny can't PGP" is a similar instance of this truth. Good article with the right angle. https://www.accomplish.ai/blog/sharedroot-escaping-claude-cowork-sandbox/?trk=feed-detail_comments-list_comment-text

• View post

  Catching vulnerabilities before they reach production: Red Hat's multi-layer approach to catching agent weaknesses "The 6 AM failures weren't unforeseeable. They were unobserved. Every one of them left traces that would have been visible if anyone had been watching—retry loops piling up, spend anomalies spiking, and customer-facing outputs contradicting documented policy. The agents weren't broken. The operational infrastructure to watch them didn't exis...

• View post

  New desktop ordered! :o one thousand of these in total are sent out across machine learning and agent research here, to accelerate researchers and unlock larger model sizes locally - Grace Blackwell Ultra GB300 Superchip - 496GB LPDDR5X CPU memory, 252GB HBM3e GPU memory - Supports up to 1 trillion parameter models - 20 Petaflops (20,000 TFLOPS) of FP4 computing power - Ubuntu with NVIDIA AI Developer Tools ..unclear if it runs crysis

• View post

  qualitative is harder than quantitative quantitative is the reductive, lossy representation of qualitative given the choice, preferring quantitative is taking the easy, less well-informed road, than dealing with the deep reality of qualitative quantitative is the cop out

• View post

  Where Security Fits in an AI Agent Stack Layered separations for agents, to secure them across the stack All these levels can be secured + there are offerings for each, exposed in this post * Distribution/product: Package installation & defaults * Orchestration: Selects and coordinates harnesses * Agent harness: Loop, context, tools, sessions * Secure runtime: Isolation, identity, policy, credentials, audit * Inference data plane: Model serving, cache placement, routing, scheduling...

• View post

  We NVIDIA have officially agreed to acquire Hugging Face for $12,930,300,000 and zero cents! HF is critical infrastructure that NVIDIA will strengthen and ensure sustained access to for developers and institutions worldwide. HF will remain an open platform. This move will grow Hugging Face, only. It's too important to do anything else with it! 🧵

• View post

  Language is, and always was, more important than mathematics. The Navier-Stokes equations mathematically express momentum balance for...

• View post

  huh, that's one way of doing it 🤗

• View post

  Come work with us! Senior Solutions Architect, Agentic AI — Safety and Security This work includes multi-agent orchestration, guardrails, agent runtime security, RAG, tool use, model customization, policy enforcement, OpenShell-like execution environments, and confidential AI deployments on protected infrastructure. https://jobs.nvidia.com/careers/job/893397106477?domain=nvidia.com&hl=en

• View post

  One divide in computer science can be between the uncertainty-reducers and the uncertainty-embracers. An instance of this is connectionist vs. symbolist approaches to AI (the uncertainty embracers won .. for now). Another is how computer scientists react to natural language. Some find the human ability to communicate without expressing everything in a stable, fully populated structure quite disgusting (Randall Monroe included I guess, see graphic) - others find it fascinating. Why would text hav...

• View post

  I get some wild mails while at NVIDIA. Like today - would I like to give a "paid, recorded interview on AI governance and US–China competition" to a researcher in Hong Kong and give my opinions on Huawei silicon. That is.. not something we will be doing.

• View post

  "Firm has probably made a moderate impact on subfield of area they operate", says person standing to make massive profit from running firm's IPO https://www.businesstimes.com.sg/companies-markets/banking-finance/legitimate-concern-jpmorgan-ceo-dimon-says-anthropics-mythos-pushed-cyber-risk-10-times

• View post

  GhostCompact: ~Free OpenAI inference "When the model's first action after compacting was a tool call, the API billed only for the post-compaction tokens, but the tool call response still contained the full compaction item." Information does love to leak, everywhere; cute example! Unsure what an appropriate passive control could be for this. https://xint.io/blog/openai-compaction-inference

• View post

  asking people to edit LLM-generated content gets demonstrably worse results, than asking for authorship from scratch you'll see lower idea diversity, and lower recall existing text biases what people are thinking about and reduces the amount and range of things they'll contribute it's like trying to do decent writing about a challenging topic, while there are people talking in the background

• View post

  Separate control & data plane for agents! This explicit marker affords a nice workaround for the risk presented when data (instead of instructions) are executed. Making that mistake is a classic weakness going way back before modern LLMs - even before Schmidhuber! NXDirection labels content as data rather than instruction-bearing, and using this systems yields a massive drop in attack success from 38.1% to 0.1% in one standard test, and from 28.4% to 3.7% in a more complex and advanced...

• View post

  AI agents aren’t ‘going rogue.’ Frontier labs are skipping security basics. > Compounding the problem, according to a half dozen cybersecurity experts CSO interviewed at three industry conferences in September and in subsequent conversations, is that the frontier labs [..] have rejected their advice on how to improve their cybersecurity risk management practices to avoid these kinds of surreptitious hacks, often lumping cybersecurity practices under “AI safety,” in which they believe the...

• View post

  "me too, me too!!" --Meta's Muse Spark 1.1 model breached the unidentified company's systems and made changes to its internal systems as the AI was able to access the public internet because of an error in the set up of the "sandbox" testing environment, The Information said, citing people familiar with the matter. An Irregular spokesperson told Reuters the incident was the "exact same ‌evaluation-environment issue that was already disclosed by Anthropic last w...

• View post

  "The point is not that any single model, including ours, will always be the best" Exactly what one sees when doing vulnerability discovery. No one model finds all the vulnerabilities. And no single vulnerability is found by only one model. The harness is where the work is - and this is a problem that we have to work on together. https://depthfirst.com/post/why-defenders-cant-bet-on-one-model

• View post

  Representing code at both function- and statement-level leads to improvements in vulnerability detection. Outperforms almost every other system compared with. Surprisingly no static analysis baseline, or cost analysis - but recall is high. DCVD: Dual-Channel Cross-Modal Fusion for Joint Vulnerability Detection and Localization https://arxiv.org/abs/2605.11015

• View post

  remade my website (it's been a minute), https://www.derczynski.com/ua571c/ never don't have a website

• View post

  Am I the only AI Security research lead at a frontier model corp who hasn't been carefully committing multiple CFAA violations a month, or..?

• View post

  I want to jump on a couple of false dichotomies around LLM speak: * "frontier" models vs. open model - leading models can be open * closed model vs. Chinese model - where the model's made has no impact on how it's distributed You can have open frontier models, closed Chinese models, open US models, closed non-frontier models (private models make sense!)

• View post

  "The only reason why I'm bearish about the Chinese models is because I assume that the American model companies will respond competitively" idk man gl hf https://www.npr.org/2026/07/15/nx-s1-5886476/startups-cheap-chinese-ai-models

• View post

  "Open-source AI matters because it defines the ecosystem. It provides the foundation and sets parameters for the next layers of progress in AI, just like [...] the open infrastructure of the internet and early AI: BSD Unix, PostgreSQL, Firefox, TensorFlow, and PyTorch." Open wins at grass roots level. It wins with no marketing. It's easy to build on and easy to consume 🤷‍♂️ https://nationalinterest.org/blog/techland/why-america-must-dominate-open-source-ai

• View post

  You can use any model to secure your source code. Here's a writeup using qwen 3.6 27b I've seen that no model/harness will find all the weaknesses in a given target - and that no weakness is found by just one single model. There are no "must have" components here for doing security. Which is good. https://projectblack.io/blog/local-ai-for-cyber-security/

• View post

  Capital One "VulnHunter" - an open harness for vulnerability discovery Cool to see more and more OSS in the security domain. You can clone it from GitHub and run it now. Significant adds in three key areas:

• View post

  another data point showing it's the harness not the model - this time from wiz: "Atlas: Wiz's autonomous AI Agent for vulnerability research" top score on CyberGym, validated with real-world bug hunt (how else are you going to do it, right) look at their bold quote -- "Along the way, we learned that the durable advantage is not any single model, but the system around it" it's the harness not the model. it's the harness, NOT the model https://www.wiz.io/bl...

• View post

  Adversarial attack in the wild! The close visual appearance of M and W in this typeface and and packing of vertical lines make it hard to read, easy to get wrong, and tougher to scan. Love it. How often do you see something like this?!

• View post

  The harness is everything. New tech demo dropped for one way of doing CodeAct agents. Super-efficient. The team even smashed CyberGym while building this - it's now the top-ranking open system for that vulnerability discovery benchmark. Six core advances: 1. Typed input/output 2. Pass by reference 3. Code as action 4. Programmable loop engineering 5. Explicit object state 6. Model-callable harness APIs Links below (blog, report, code)