搜索“Agent”

We’ve built an AI agent that acts as a secondary expert for a given domain, making deep specialist knowledge readily available and preserved for anyone in an organization to access, share, and build upon; This is not a typical domain-specific agent. Its novelty comes from integrating two layers: A structured, auditable knowledge architecture separates what the agent knows from how it reasons; A self-improvement loop then compiles expert feedback into verified, regression-tested updates without m

匹配字段正文

cases; Developers annotate Rust source code directly with preconditions and postconditions using Rust-like syntax, enabling fast feedback loops (under one second) and allowing AI agents to assist in proof generation; Verus enables mathematical verification of Rust's "unsafe" code blocks and concurrent code with custom locking schemes, re-establishing machine-checked safety guarantees for performance-critical implementations like AWS's Nitro Isolation Engine; Amazon uses Verus to prove correctne

匹配字段正文

similar role for quadruped robots — though as I’ll discuss later, Unitree’s robots now face legal restrictions in the United States. Sponsor message Tasklet — Run your business on agents AI agents are useful when they stop chatting and start doing. Tasklet lets you build persistent agents that work across Gmail, Slack, HubSpot, Zendesk, Linear, and thousands of other tools. Describe a workflow in plain English, choose what the agent may do on its own, and require approval for sensitive steps. Ag

匹配字段正文

ecific deployment point defined by the model, workload, SLO, serving topology, runtime version, and accelerator platform. The same patch may improve one point and regress another. Agent-assisted exploration can produce more environments, experiments, measurements, and rejected paths, all of which must remain reproducible. The first requirement is therefore reliable execution. Reproducing a deployment point requires more than model capability: tools, environments, context, memory, Verification, a

匹配字段正文

g a long tail of UI regressions. As our team grows, we wanted to create a stronger foundation for how UI is composed. Defining these boundaries is becoming especially important as agents contribute more and more to our codebase. Another more practical reason to migrate was performance. We care about keeping Linear fast and CSS-in-JS adds work during rendering to generate styles and inject CSS rules. The final straw was the performance regression we saw after upgrading to React 18 that introduced

匹配字段正文

boot the machine when it hardlocked the host kernel. If it wasn’t clear before, I will state it plainly: you can no longer assume a mere VM will contain a sufficiently advanced AI agent. To use a 2010s term of art, you should treat such agents as an advanced persistent threat. Let’s imagine a common situation: you want to run an AI agent, but need its full capabilities, including the ability to download, build and debug software; do research; sift through dependencies; set up fuzzers; triage vul

匹配字段正文

dpoints, dedicated deployments, and training support. The model provides a 1M-token context window with native tool calling and preserved reasoning history, targeting long-horizon agentic workloads. We ran DeepSeek V4 Pro 0813 through the same eval suite we used for Kimi K3 in July . We evaluated Deepseek against Kimi K3 and Fable 5 across four benchmark families: SWE-Bench Verified, LiveCodeBench v6, Aider Polyglot v1, and Terminal-Bench 2.1. Deepseek V4 Pro is the strongest single model in thi

匹配字段正文

1,507 real vulnerabilities across 188 widely used open-source projects, drawn from Google’s OSS-Fuzz campaign and the ARVO reproducible-vulnerability archive. Each task hands the agent a short description and a real vulnerable codebase. The agent has to locate the flaw, build an input that actually triggers it, and then patch it. Grading is by execution, not by a judge model. The proof-of-concept either crashes the target or it does not. The structure of Cybergym is agentic, long-horizon, and u

匹配字段正文

Engineering Browser agent benchmarks are saturating. This is how we build RL environments from real user traffic against the real web. Evaluations 1. Filtering Down the Tasks; 2. Suitability; 2.1 Hand labelling; 2.2 Aligning a labelling agent; 3. Personal Data Replacement; 4. Feasibility and Difficulty; 5. The Problem with Judgement; 6. Difficulty Estimation; 7. Rubric Construction; 7.1 Why iterative rubrics do not converge; 7.2 The parallel rubrics we use instead; 7.3 Properties of a good rubri

匹配字段正文

ice as fast per rollout. GLM-5.3 ties pass@2 and wins pass@4 (87.6% against 85.8%) at half the cost per rollout, so it is the stronger value pick for high-volume or retry-tolerant agent work. In our run, GLM-5.3 cost \$3.99 per rollout against \$8.37 for GPT-5.6 Sol at max effort, about 2.1x lower. Measured per solved task, GLM-5.3 returned 17 solves per \$100 against Sol's 9, roughly twice the solved work per dollar. They split the board. GLM-5.3 takes JavaScript (90 against 75) and Rust (70 ag

匹配字段正文

and is much faster per rollout. DeepSeek V4 Pro 0813 wins pass@4 (88.5% vs 85.8%) at 35x lower cost per rollout, so it is the stronger value pick for high-volume or retry-tolerant agent work. In our run, DeepSeek V4 Pro 0813 cost \$0.24 per rollout versus \$8.37 for GPT-5.6 Sol at max effort, roughly 35x cheaper. Measured per solved task, Pro returned 260 solves per \$100 against Sol's 9, about 30x the solved work per dollar. GPT-5.6 Sol leads Python (74), Go (79), TypeScript (66), and JavaScrip

匹配字段正文

an applied end-to-end LLM project including evaluation, training, and local deployment for real-world use. The outcome of this is an AI-detector API that can be used by humans and agents, and a user-friendly UI. Method overview Here, we are going to develop a method similar to Pangram models, which, as far as I know, are behind Substack AI detection feature. I wrote a short article about AI-text detection a while back in 2023: What Are the Different Approaches for Detecting Content Generated by

匹配字段正文

cky decisions: take the native search some labs build in, or wire up a third-party engine like Exa, Parallel, or Perplexity? Is one search enough, and if not how long do I let the agent keep searching? Are more search turns worth the quality they buy? We built live leaderboards to help you decide the best search configuration with data. See the data on our new Benchmarks page. We benchmark all the combinations to find strengths and weaknesses When setting up a search request, you have four decis

匹配字段正文

You wouldn't let Claude Code cat your AWS credentials or Kubernetes config on request, would you? A browser-based game designed to test humans' ability to safely approve AI coding agent requests suggests humans in the loop aren't as good at spotting dangerous commands as one might hope, with players approving roughly one in three malicious requests on average. The results also suggest that repeatedly having to approve an agent's actions can lead to sloppy decisions. It’s a quick, simple game on

匹配字段正文

I / Cybersecurity English 中文 The NVIDIA AI Red Team shares how access controls, sandboxing, network egress restrictions, and secret management can reduce the risks of deploying AI agents at enterprise scale. AI-Generated Summary The NVIDIA AI Red Team assessed multiple AI agents over the past six months and identified four recurring failure modes: lack of access control, tools enabling arbitrary code execution, missing network egress controls, and secrets exposed in plaintext. Prompt-based guard

匹配字段正文

Inside Figma; Engineering; Security Our security team built an AI agent that triages alerts, conducts forensic investigations, queries our security data lake, writes code to fix issues—and remembers what it learns. Here's how we cut alert time-to-resolution by 71% and fundamentally changed how our on-call engineers work. Share How we secure Figma’s internal systems with agents Illustrations by Jimmy Simpson When it comes to protecting our internal platforms, the threats we defend against on Figm

匹配字段正文

Trail of Bits July 28, 2026 open-source , vulnerabilities , machine-learning Page content 1. Let Codex write the goal 2. Define the outcome, not the path 3. Assign one outcome per agent Where human judgment is needed Codex’s /goal feature amplifies bug hunting, but getting good results requires the right prompt, the right scope, and the right number of outcomes per run. For Patch the Planet , our joint initiative with OpenAI to find and fix bugs in open-source software, we pointed Codex at some

匹配字段正文

sounds to humans WER is a necessary metric for many use cases, especially those involving alphanumerics that are not words. But WER is rarely sufficient for enterprise-grade voice agents. 3. Contextual correctness is another challenge Many errors require semantic disambiguation to even define correctness: “read” (present vs past); “lead” (metal vs verb); “Polish” (nationality vs verb); “bass” (fish vs instrument); “live” (verb vs adjective); years like “2026” (twenty-twenty-six vs two thousand t

匹配字段正文

(pass@1 72.7% vs 68.5%) and solves more tasks four-for-four. Kimi K3 wins pass@2 and pass@4 and costs far less, so it is the stronger value pick for high-volume or retry-tolerant agent work. For most teams the best answer is to route between them. In our run, Kimi K3 cost \$4.65 per rollout versus \$8.37 for GPT-5.6 Sol, a little over half the price. Measured per solved task, Kimi K3 returned 14.7 solves per \$100 against Sol's 5.3, about 2.8x the work per dollar. Yes, if you can verify results

匹配字段正文

We removed over 80% of Claude Code's system prompt for more advanced models. How to apply the lessons we learned to your own context engineering in Claude Code and with your own agents. I’ve written previously about how to best prompt the newest generation of Claude 5 models and work with them iteratively to discover what you want to build. But when you send a message to Claude, the prompt is only a small part of the context it gets. Much of your context is assembled from your system prompt, Ski

匹配字段正文