Reward Hacking과 에이전트 보안, 로컬 모델 실행과 추론 인프라

New research: Training a Misaligned Reward Seeker What produces severe misalignment? We’ve long been concerned that cheating during training—otherwise known as reward-hacking—might teach a model to pursue rewards by any means available. To study this at scale, we trained an Opus-sized model on 80 production environments we knew to be hackable. In simulated evals, it engaged in unauthorized cyberattacks, tampered with its reward, and tried to evade safety monitoring. Read more: http:// alignment.anthropic.com/2026/reward-se eker …

We’re sharing an update on our alignment and security efforts. In July, we reported three incidents in which Claude models, running without safeguards in cybersecurity evaluations, gained unauthorized access to real systems. In a new post, we describe: 1. How we’ve secured our evaluation and training environments, and practices we've asked external partners to adopt when testing pre-release models without cyber safeguards 2. An update on our alignment assessment 3. New research on how reward hacking during training shapes model behavior, why we think our work this spring kept these incidents from being more severe, and why gaps in that work may have contributed to them 4. How we hardened our security practices earlier this year to prepare for Mythos-class models Read more:

Elon Musk

@elonmusk

2일 전

Grok @Bot upgrades

Elon Musk

@elonmusk

2일 전

Grok @Bot

Open-sourced a peanut-sized OCR that parses entire 100-page PDFs in one shot & Runs locally. - Only 3B params. - Already 3M downloads Every other OCR tool chops your doc into pages and loses the thread. this one reads the whole thing in a single pass. - One-shot "long-horizon" parsing (32K context window) - Multilingual, out of the box - 93% on the standard parsing benchmark (+6 over baseline) - 0.11 error rate past 40 pages - Runs 100% locally on your own hardware - Works with Transformers, vLLM, SGLang, Docker, Ollama, llama.cpp Traditional cloud OCR (Textract, Google Vision, Azure Doc Intelligence) costs $1.50-$15 per 1,000 pages. This runs on your machine. https:// x.com/0x0SojalSec/st atus/2075022982137974985/video/1 …

There’s generally a massive difference in quality between specialized OCR providers, “simple” open-weight OCR models, and free/OSS solutions. Specialized OCR models (including LlamaParse) solve for the long-tail of complex documents, and make sure that everything is digitalized properly with lower hallucinations. They typically use posttrained VLMs to cover a wide range of real-world docs. They have tuned bounding boxes and annotations for each section, letting agents trace citations back to the source. They also usually come with additional endpoints like extraction and splitting. Open-weight VLMs (e.g. Paddle, MinerU, UnlimitedOCR) are reasonable over relatively simple documents like text and tables and can do basic visual reasoning. They can seem somewhat cheap to host but can be unreliable in quality. Free OSS libs (including liteparse) are meant to be universally accessible, fast text extractors. They’re not meant to do any sort of visual reasoning, so won’t perform any linearization, or reasoning over complex tables, or OCR over non-native docs. AI agents like Claude will by default use these tools to do a light pass over documents. But I would caution using these for retrieval, because they will drop entire sections that are not digitalized. At this point we’ve benchmarked over 92 tools on ParseBench. Come check it out! https:// parsebench.ai

How fast is the Claude Code plugin ecosystem actually growing? Plugin-touching commit activity grew 8.8x in the six months after the initial launch. Researchers studied 1,926 repositories hosting plugin marketplaces, covering 8,351 plugins across 2,018 marketplaces and 77,773 commits. These are maintained artifacts rather than write-once files. Feature commits run at 39.6% against 17.2% for conventional open source. Software engineering tasks account for 61.3% of all plugins, and Claude co-authors 34.9% of every commit in the dataset. Natural-language instruction files and their implementation scripts co-evolve at above-chance rates, and 78% of those co-changes are functionally coupled. Prose and code have become one versioned unit, a maintenance dependency with no analogue in traditional software. Paper: https:// arxiv.org/abs/2608.28497 Chat with Paper: https:// academy.dair.ai/papers/on-the- maintenance-and-co-evolution-of-agent-plugins-an-empirical-study-of-claud-2608.28497 …

// ContextLeak in AI Agents // The whole attack surface here is a tool name and a tool description. Stealing an LLM agent's runtime context, meaning the user prompt, the execution trajectory and the tool list, needs three things to line up. The agent has to pick the malicious tool, it has to pass its context in as arguments, and the tool has to forward that anywhere the attacker wants. Existing work covers the first and third conditions and leaves the second one mostly alone. ContextLeak targets the middle step. Researchers at Duke use an attack LLM to generate the malicious tool's name and description, then fine-tune that LLM with reinforcement learning on a set of shadow users with diverse simulated agent contexts. The reward functions are built specifically for the exfiltration objective. It remains highly effective when the shadow contexts differ substantially from the victim's, and it outperforms existing malicious-tool attacks adapted to this setting. Paper: https:// arxiv.org/abs/2608.27800 Chat with Paper: https:// academy.dair.ai/papers/context leak-exfiltrating-llm-agent-context-via-malicious-tools-2608.27800 …

Qwen3.8-Flash-Next by @Alibaba_Qwen has landed in Agent Arena, ranking #7 among open models (#24 overall) with +2.4% net improvement across 8.7K+ real-world agentic sessions! Among open models, it sits just behind DeepSeek V4 Flash (High) at #6 (+3% net improvement) and two spots behind GLM-5.3 (Max) at #5 (+3.8%). By signal, Qwen3.8-Flash-Next stands out in delivering an explicit response from the community on task completion (Confirmed Success at +12.3%), ranking #5 among open models (#7 overall). More detail on its performance by signal below. Qwen3.8-Flash-Next outperforms Qwen3.8-27B in overall net improvement (+2.4% vs. +1.5%) and Confirmed Success (+12.3% vs. +7.2%). Qwen3.8 Max remains ahead overall at +6%, with +10.8% Confirmed Success. Congrats to the @Alibaba_Qwen team on this contribution to the open ecosystem!

Teams across Google have been building with (and loving) Gemini 3.7 Flash. We’re seeing: Real-time website generators 3D physics simulators Interactive webcam tools Personalized field guides Take a look at how Googlers are using 3.7 Flash across @GoogleAIStudio , @Antigravity , and @GeminiApp Spark

You can now run an infinite AI TV show that listens to Twitch chat and generates the next scene live. - Real-time video gen (LTX) - Twitch chat to prompt - Seamless looping with context - RTMP to Twitch - Live dashboard Open source, Infinite TV - https:// github.com/alex-remade/in finite-tv …

Locally 4090 MiniMax Fast H3 can generate an entire TV series indefinitely on your machine. Using the 4090 at 1344×768 resolution, generating 15 seconds of content only takes 151 seconds, which is very fast.

Keeping every model loaded wastes GPU capacity. Swapping them only helps if the swap is fast. Semi-Persistence keeps weights in a pinned CPU pool and treats the GPU copy as disposable. 2B to 397B params: sleep/wake cycle 5.6x to 19.9x faster. Sub-second on single GPU. Now open source from Snowflake AI Research: https:// bit.ly/45UPGRt

for september, we’re cutting the price of Dedicated Inference on H100s from $5.49/hr to $3.99/hr new + existing deployments get the lower price automatically deploy gemma 4, qwen3/3.5, gpt-oss, llama, nemotron 3.5 lightning models, or bring your own lora for a fine-tuned model