Show HN: Claw Patrol, a security firewall for agents
Claw Patrol is a security firewall designed for AI agents. The project, presented on Hacker News, aims to control and secure the actions of autonomous agents.
2731 articles
Claw Patrol is a security firewall designed for AI agents. The project, presented on Hacker News, aims to control and secure the actions of autonomous agents.
TTS benchmark revamped with objective rating system and blind voting. 46 models evaluated through live ELO arena. Hugging Face interface and GitHub repo available for testing local TTS models.
Cohere releases North Mini Code, its first model designed for developers. The model is optimized for code generation and completion, with multilingual support and integration into the Hugging Face ecosystem.
A German regional court rules Google is directly liable for AI Overviews content. The court rejects limited liability protections that apply to traditional search engines. Google's AI falsely linked two publishers to fraud with claims absent from cited sources.
A judge cancelled a trial after discovering both parties used AI to prepare their cases. The magistrate removed the attorneys from the case, raising questions about unsupervised AI tool use in legal proceedings.
Google DeepMind releases Gemini 3.5 Live Translate, a near real-time natural speech translation system integrated into Google AI Studio, Google Translate, and Google Meet. Translation maintains fluidity and naturalness of original speech.
Apple Intelligence enables Siri to understand and analyze iPhone screen content. The feature transforms the interface into an intelligent interaction ground for contextual actions.
Study comparing LLM performance against classical hyperparameter optimization algorithms. Results show respective strengths and limitations of each approach for model tuning.
Warning about OpenCode Go/Zen: users report inability to delete accounts. Multiple GitHub issues remain unanswered for extended periods. Developers vaguely respond they will 'probably' add account deletion functionality.
Practical guide to building a basic AI agent capable of planning long tasks. Covers fundamental concepts and implementation of multi-step planning systems.
Chinese makers are producing single-slot, half-height V100 PCIe cards with NVLink. Custom PCB with soldered core, passive cooling, 75W default or 300W alternative version. 16 cm × 7.5 cm. Expected ~$220 USD (16 GB), 32 GB version coming. Bilibili video, not yet for sale.
A user deployed an arbitrage bot on Polymarket and shares actual performance results. The article details returns, costs, and challenges encountered while running an automated strategy on the prediction market.
Google DeepMind releases Gemma 4 12B, a unified encoder-free multimodal model. The model processes text, images, and video in a single architecture, optimized for on-device inference.
Text-to-CAD generation method using LLMs to produce controllable 3D models faithful to instructions. Combines natural language processing with 3D geometry to convert textual descriptions into usable CAD files.
Google DeepMind announces investment in European robotics to develop advanced autonomous systems. The initiative aims to strengthen research capabilities and practical applications of robotics across the continent.
China plans to invest $295 billion in a nationwide AI data center network over five years. At least 80 percent of components would come from domestic suppliers like Huawei, excluding US vendors. Taiwan is considering criminalizing AI chip smuggling to China.
OpenAI enhances ChatGPT's memory with a system connecting past conversations to current needs. This feature becomes available to free users.
Apple announced CoreAI at WWDC, an on-device inference engine to replace CoreML. Supports models up to 20B parameters via Python conversion. Performance vs MLX/llama.cpp not yet documented.
Article exploring how AI agents use advanced search techniques beyond basic approaches. Analyzes architecture and capabilities of modern agent systems to optimize information search and retrieval.
Rust-native CPU-only implementation of LFM2.5-8B-A1B with tool use callbacks. Decode ~37 tokens/s on Ryzen 7950X, prefill not yet optimized. Memory footprint ~7GB, runs on 16GB RAM. Published as cargo crate.
A post-trained model that performs penetration testing instead of refusing code analysis. Approach contrasting with standard LLM safety guardrails.
Nextdoor engineers use Codex with GPT-5.5 to investigate hard-to-reproduce issues, build across platforms, and focus on product outcomes.
Decentralized prototype using local embeddings (EmbeddingGemma-300M) to replace central indexes. Devices communicate peer-to-peer, rank content by semantic distance (cosine similarity) without server or global ranking. Proposed extension to AI agents discovering each other's needs/offers through semantic proximity.
Apple rolls out a security feature that automatically replaces compromised passwords before they can be exploited. The functionality detects breaches and initiates password changes without user intervention.
Agent-skills: open-source framework to equip AI coding agents with production-grade engineering capabilities. GitHub repository aiming to standardize coding agent competencies.
GitHub repository collecting system prompts and internal models from 25+ AI tools (Claude Code, Cursor, Devin AI, Perplexity, Replit, v0, etc.). Includes open-source alternatives. Resource to reverse-engineer system instructions of popular assistants.
Ataraxy-Labs/sem: semantic version control on top of git with entity-level diffs, blame, and impact analysis. Supports 26 languages via tree-sitter. Built for coding agents.
Cube Core is an open-source semantic layer for AI, BI and embedded analytics. The project is gaining traction on GitHub Trending.
Chroma is a vector search infrastructure for AI applications. The trending GitHub project provides storage and querying tools for embeddings to support RAG and language model-based systems.
Asm is a universal skill manager for AI coding agents. The GitHub project provides infrastructure to orchestrate and manage capabilities of autonomous coding agents.
DesktopCommanderMCP is an MCP server for Claude providing terminal control, file system search, and diff-based file editing capabilities.
ARIS (Auto-Research-In-Sleep): lightweight Markdown-based framework for autonomous ML research. Cross-model review loops, idea discovery, and experiment automation. Works with Claude Code, Codex, OpenClaw, and any LLM agent.
GitHub Action powered by Claude that automatically analyzes code changes to detect security vulnerabilities.
Reddit discussion on real-world adoption of privacy-preserving ML techniques (differential privacy, federated learning, on-device inference) in production systems. Active research literature noted, but actual industrial deployment questioned; explores engineering challenges, performance/cost impact, and use cases.
Nvidia bets on AI PC demand beyond niche users, but relies on unproven use cases and uncertain mass-market adoption. The strategy faces validation challenges in mainstream consumer segments.
Apple unveils a rebuilt Siri at WWDC 2026, powered by foundation models developed with Google. For complex queries, the assistant leverages Nvidia GPUs.
Jetson Orin NX build for compact LLM server with benchmarking. Gemma 4 26B quantized Q2_K_XL achieves 66K context window, 14.65 tok/s at 8K context and 10.21 tok/s at 60K context. Supports multiple tool calls for Hermes agents.
An AI agent built a 3D Paris gallery by chaining two Hugging Face Spaces. The system automatically orchestrated image generation and 3D environment creation without manual intervention.
LG Group strengthens its partnership with NVIDIA to build an AI Factory infrastructure designed to accelerate its artificial intelligence projects.
OpenAI abandons its goal of full automation by 2028, favoring human-machine collaboration instead. Altman and Pachocki call for an international body capable of slowing frontier model development if needed.
The Weather Company migrates infrastructure to Vercel and v0, serving 350 million monthly active users. Redesign cuts design-to-publication time from days to hours and increases velocity by 80%. 2.2 billion coordinates calculated every 15 minutes.
TinySearch v0.2.0 (first stable beta) switches from DuckDuckGo to SearXNG as default search backend. This lightweight MCP/FastAPI tool for small local LLMs crawls pages, chunks and reranks them to provide compressed context (max 8k tokens) to agents, avoiding prompt bloat. Tested with Qwen 3.5-9B.
OpenAI has confidentially filed an S-1 registration with the SEC, the first formal step toward an IPO. The company calls it "a complicated set of tradeoffs" with no set timeline. Anthropic's recent IPO filing adds competitive pressure.
OpenAI prepares its IPO following a confirmed confidential filing by the ChatGPT editor, days after Anthropic's similar announcement.
A researcher fine-tuned a language model to generate content optimized for ADHD brains by maximizing dopaminergic engagement. The approach combines fine-tuning on curated data and evaluation via behavioral metrics.
INSEE reports French growth at +0.9% acquired by mid-2026. AI's impact on economic activity remains modest despite expectations.
Uber and Wayve pursue a fleet/software separation model for robotaxis in London (June 2026), while Waymo favors vertical integration. The two strategies diverge on value chain control.
Red Hat and NVIDIA announce joint infrastructure for AI agents at Red Hat Summit 2026. The partnership strengthens both companies' agentic AI capabilities.
A researcher tests Gemma 4 31B on understanding complex, niche academic code. Gemma 4 31B outperforms Qwen 3.6 (27B and 35B) and matches Claude Opus 4.7 in grasping inter-component dependencies. Qwen 3.6 shows excessive eagerness but spots a local improvement both other models miss.
Microsoft's open source tools were compromised in a supply chain attack targeting AI developers' credentials. Attackers gained access to password repositories and developer accounts.
Latent Space introduces FrontierCode, a benchmark for evaluating code quality from AI systems beyond surface-level metrics. The tool measures robustness and reliability of solutions rather than mere functionality.
Quasar-Preview by silx-ai, an open-source model with 5M token context length, is now available on Hugging Face. No technical details or benchmarks provided in the post.
Benchmark comparison of Gemma 4 26B across 4-bit, 6-bit, and QAT 8-bit quantizations on MLX. Tests: 50 MMLU_PRO questions and 100 HumanEval. 6-bit achieves 58% MMLU_PRO and 98% HumanEval; 4-bit 56% and 90%; QAT 8-bit 52% with incomplete HumanEval results.
LFNO (Laplace-Fourier Neural Operator) combines spectral advantages of Laplace and Fourier operators via explicit decomposition into transient and steady-state regimes. Evaluated on 9 benchmarks (3 ODE systems, 6 PDE systems including Navier-Stokes), LFNO outperforms existing operators on ODEs and matches FNO performance on PDEs.
New evaluation method for social AI agents: an evaluator agent actively interacts with the target agent to generate situations testing specific social criteria (conflict handling, etc.). Tested on 32 criteria in a life-simulation environment, improves coverage and agreement with human labels compared to passive methods.
IntentPOI, a two-stage LLM-based framework, predicts next Point-of-Interest by first inferring user intention from historical mobility, peer behaviors, and temporal context, then selecting POIs aligned with that intention. Outperforms 11 baselines on three real-world datasets.
PPV (Propagational Proxy Voting) outperforms majority voting on MMLU-Pro (+1.5 pp, +2.24 pp on non-trivial subset, p~1.0e-14). This unsupervised aggregator leverages letter entropy and reasoning geometry to weight 128 sampled generations partitioned into 16 groups, requiring no gold labels or auxiliary training.
New industrial dataset MMIO (80K+ samples, 6 super-categories, 18 subcategories) for zero-shot defect detection. RTVP method (Refined Text-Visual Prompt) based on Mobile-SAM with expert-guided domain adaptation. Achieves 42.2% AP zero-shot and 24.7% AP in closed scenes.
AI-MASLD, a stress-audit framework, evaluates 7 medical LLMs on 240 clinical cases with narrative perturbations. All perform well at baseline but diverge under realistic stress. Quantized models hide functional collapse; medical fine-tuning degrades logical stability and fairness. An open-weight model matches or exceeds proprietary alternatives on all safety dimensions.
EditSR introduces a two-layer framework for neural symbolic regression. First layer generates expressions via autoregressive decoding, second layer (Rectifier) corrects syntactic errors through step-by-step editing. Rectifier is pretrained to maintain efficiency without restarting global search. Significant gains on complex expressions.
ThinkBooster is a unified framework for test-time compute (TTC) scaling of LLM reasoning. It includes a modular Python library, a benchmark evaluating performance and computational efficiency, and an OpenAI-compatible proxy service. Results on mathematical and coding tasks demonstrate performance-compute trade-offs of TTC strategies.
An agent-to-agent communication protocol (RCP) automates exchanges between regulators and applicants in advanced nuclear reactor review. Tested on 1,236 NRC documents, it reduces costs by 50-77% (21-44M USD vs 89M USD) and timelines by 65% (15 months vs 42 months). Applicable to other regulated sectors, potential savings reach 210-330 billion USD/year.
New paradigm for modular AI systems built from bottom-up contributions of specialized small models. Participatory systems outperform monolithic LLMs by 15.4% across 15 tasks (reasoning, factuality) and solve 15% of problems where all individual models fail.
Prithvi-EO-2.0, a geospatial foundation model, tested on 19 flood events (2017-2025) across 6 continents. Accuracy varies by land cover: cropland 52% IoU, tree cover 4%. Riverine detection strong (F1=0.69). 23 failure modes identified; pipeline engineering dominates initial errors over model capacity.
PathoSage is a three-stage framework for multimodal reasoning in computational pathology. It explicitly separates knowledge retrieval, evidence collection, and evidence adjudication via Structured Evidence Deliberation. A training-free Beta-Bernoulli experience system models tool reliability to reduce hallucinations and anchoring bias.
MemToolAgent improves LLM agent tool use through structured memory management. The framework extracts past experiences into memory entries, dynamically retrieves relevant ones, and generates reflections from user feedback. Achieves 29%, 80%, and 17% relative improvements on WorkBench, NESTFUL, and PEToolBench without fine-tuning.
Theoretical paper on chatbots' cognitive limitations in problem-solving. Authors argue LLMs encode artificial 'metaphorical problem propagations' from training data, unable to replicate human understanding. Conclusion: further LLM development will not enable chatbots to become human-equivalent thinking partners.
IRSL integrates Item Response Theory (IRT) into LM scaling laws, reducing complexity from O(M×N) to O(M+N). Validated on 6,612 checkpoints and 37,682 questions, Beta-IRT achieves comparable accuracy with 99.9% fewer questions (50 per benchmark).
Benchmarking study showing that interpretability methods (IML) applied to genomic models often yield contradictory explanations and fail to localize known regulatory motifs. Authors propose a rigorous evaluation framework inspired by clinical trials to replace current anecdotal validation practices.
New 'Contribution Weights' metric for analyzing transformers beyond attention weights. Incorporates value vector magnitude and directional alignment. Outperforms attention-based metrics for identifying critical tokens. Reveals attention sinks play active information suppression role, stabilizing representations against semantic drift.
LLM-based LEGO assembly generation suffers from PhysHack: physically valid but geometrically misaligned structures. Authors propose PVPO, a sample-efficient RL method coupling physical feasibility with voxel-space geometric rewards. Results: improved semantic alignment, structural stability, and calibration across model backbones.
DiffOR introduces a novel paradigm for Ordinal Regression as Continuous Generative task. The framework leverages diffusion models to recover continuous ordinal values via iterative denoising, with a Dual-Decoupling Strategy (Multi-scale Increment Aggregation and Dynamic Denoising Perception) to preserve ordinal topology. Validated on 12 benchmarks across four domains.
A post-hoc method reduces fine-tuning bias by truncating the tail of the SVD decomposition of weight updates (ΔW). Tested on 3 models (0.5B–7B) and 4 benchmarks, it decreases performance gaps on underrepresented groups up to 5× (CivilComments) with <2pp accuracy loss, requiring no retraining or group labels.
Theoretical paper proving that optimal sequential filter ordering minimizes total expected cost by sorting filters by increasing cost-to-rejection-probability ratio. Monte Carlo simulations confirm strict dominance over common heuristics across all runs.
MST-Direct scaled to large-scale multivariate and conditional geostatistical simulation via Sinkhorn optimal transport. Addresses scalability (O(nC) memory), multivariate extension, and kriging-based conditioning. Validated on 6-variate heteroscedastic distribution, grids 200×200 and 100×100 with 200 hard-data samples. Reproduces joint distribution exactly versus PPMT approximation.
Diffusion language models (DLMs) use bidirectional attention, invalidating standard KV caching techniques for shared prefixes. Researchers propose bicache, a method that dynamically identifies safe layer depth for reusing shared prefix KVs. Result: 36–98% throughput improvement without accuracy collapse.
STARIXNet is a lightweight neural network for real-time resource allocation in cloud platforms. It captures spatio-temporal relationships among multiple system metrics (seasonality, trend, auto-regression, exogenous variables) and prioritizes service stability over forecast accuracy. Deployed at Walmart, it achieves 10-50% cost savings.
RL4F is an open-source offline reinforcement learning benchmark for plasma control in nuclear fusion. Built on historical data from the DIII-D tokamak, it evaluates imitation learning and offline RL methods on four multi-actuator tracking tasks (rotation, density, temperature, pressure). Offline model-based RL methods achieve best average performance.
Passenger queue forecasting framework for airport departure gates and security checkpoints. Transformer-based architecture captures temporal dependencies and inter-facility correlations from queue lengths, waiting times, and passenger throughput. Accurate predictions up to two hours ahead.
HASA proposes subnet allocation for heterogeneous federated learning with resource and data constraints. The method assigns subnet widths based on heterogeneity scores computed from local data while enforcing fixed compute budgets. On next-word prediction (7 clients), HASA improves mean accuracy from 13.82% to 14.32% and strengthens worst-client performance.
Researchers propose a GNN framework to classify finite groups by solvability using graph representations including Cayley graphs. The model learns to distinguish solvable from non-solvable groups using structural graph information alone, evaluated on groups outside the training set.
ScaleSweep optimizes NVFP4 quantization (hardware-supported FP4 4-bit format) of LLMs through block scale candidate sweeping. Theoretically bounded to reduce search space, the method preserves >93% full-precision performance on Llama and Qwen under end-to-end quantization (weights, activations, KV cache, query states).
Query Lens extends Logit Lens to interpret sparse autoencoder features by jointly analyzing encoder-side keys and decoder-side values. The method captures indirect effects through downstream modules, revealing coherent token signatures for features opaque under Logit Lens. Hypothesis: downstream modules read features through layer-specific subspaces.
Structured pruning framework using multi-armed bandit (MAB) algorithms to remove complete neurons from deep neural networks. Evaluates UCB1, Thompson Sampling, Epsilon-Greedy and other policies on classification, regression and deep learning tasks. UCB1 and Thompson Sampling outperform magnitude-based pruning and unpruned models.
Study applying Random Forest Recursive Feature Elimination to Nigerian household survey data (2018/19) to identify minimal predictors of poverty status, welfare quintiles, and inequality. RF-RFE achieves 90% accuracy for poverty with 5 income variables, 80% for seasonal quintiles. ML methods reduce data requirements while preserving distributional information for poverty and inequality monitoring.
LEAF is a retrospective tree-based RL method for speech-aware LLM post-training that improves credit assignment by grouping responses by shared prefixes and assigning span-level advantages. It outperforms GRPO on speech question-answering and speech translation benchmarks without online branching or additional decoding.
QDSP is a structured learning framework to predict mortality or cerebral palsy in very low birth weight infants. On a cohort of 51 patients, it achieves 92% accuracy and 0.9714 AUC, outperforming XGBoost, TabNet, and TabPFN. Interpretability via SHAP identifies clinically relevant predictors including cystic periventricular leukomalacia.
SRT introduces a super-resolution framework for time series using disentangled rectified flow. The method decomposes input into trend and seasonal components, aligns them via implicit neural representation, and employs cross-resolution attention for high-resolution detail generation. SRT-large, a pre-trained scaled version, demonstrates zero-shot capabilities across 9 public datasets.
MetaEvo introduces a two-stage framework for continual LLM agent evolution. It combines preference-based optimization to enhance principle abstraction, then accumulates and reuses these principles in a modular agent architecture. Results on reasoning benchmarks show consistent improvements across iterations without performance plateaus.
Theoretical study formalizes data propagation through Transformers as a nonlinear control system. For mean-field Transformer with self-attention and affine feed-forward layers, Gaussian distributions remain exactly Gaussian. This reduces dynamics to finite-dimensional bilinear control system governing mean and covariance evolution, connecting Transformer expressivity to Riccati-type equations.
Topological framework for comparing trained GNNs by mapping Stochastic Block Models onto the n-dimensional sphere. Leverages compactness of graphon space, Frieze-Kannan weak regularity lemma, and Lipschitz continuity of MPNNs. Produces low-dimensional fingerprint for transfer-learning candidate retrieval without retraining.
Pre-training data mixture experiments fail to scale because repetition rates of high-quality data shift with training budget. A subsampling procedure matching target repetition rates recovers optimal mixtures using only 1/16 of target tokens (757M model), reducing error from 0.75 to 0.05 compared to uncontrolled baselines.
UNIQ introduces conformal calibration for adaptive conservatism in offline reinforcement learning. Built on IQL, the method uses a multi-expectile ensemble and split conformal prediction for distribution-free uncertainty estimation, dynamically adjusting penalties based on local data coverage. On D4RL MuJoCo, UNIQ outperforms IQL with 10× lower memory than EDAC.
ResearchClawBench benchmarks autonomous scientific research agents across 40 tasks spanning 10 scientific domains. Claude Code scores 21.5/100, Claude-Opus 20.7/100. Failures concentrate in experimental protocol mismatch, evidence mismatch, and missing scientific core.
Study of 21 LLM routing methods across 5 benchmarks reveals a routing plateau: most converge to similar accuracy far below oracle performance. The bottleneck is predictability—routers learn global averaged trends rather than query-specific signals. Larger training datasets, stronger encoders, and end-to-end fine-tuning improve routing accuracy.
arXiv paper introducing hybrid FT-Transformer + XGBoost architecture with calibration-aware stacking for customer churn prediction on structured data. Achieves 62.10% F1 and 0.861 AUC-ROC on public bank dataset, outperforming MLP baseline by 3.37 F1 points. Handles class imbalance using class-weighted loss without synthetic oversampling.
Theoretical paper proposing kernel contracts to bound divergence between training and inference kernels in post-training. Framework specifying acceptable gaps in finite precision with numerical, statistical, and routing clauses. Derives bounds from logit drift to total-variation distance and applies to RL policy-gradient bias.
CARTOGRAPH is a verification layer for autonomous AI scientists, combining experiment steering, explicit ambiguity closure, and library inadequacy detection. Across five testbeds, CARTOGRAPH-A outperforms raw projection (129W/0T/15L, p<10^-21). The refuse guard correctly flags all 4/4 claims later marked inconclusive in manual A-Lab reanalysis.
Mathematical framework (HEF) modeling emergence as phase transition in mechanism landscape. Empirical study across 111 grokking experiments in modular arithmetic transformers: convergence to 0.9745±0.014 regardless of initialization, weight norm peaks before grokking in 92% of runs, accuracy curves collapse onto tanh kink (R²=0.93).
SPIN is a decentralized coordination framework for multi-agent swarms on edge platforms. It models swarm topologies as compressed tensor networks, reducing complexity from O(n^m) to O(m·n·χ²). A hybrid neuro-symbolic pipeline combines offline pre-trained neural coordination encoders with zero-shot reweighting filters based on the Radon-Nikodým derivative.