Boring AI News

Anthropic says its Claude Mythos Preview identified thousands of zero-day vulnerabilities across major operating systems and browsers. Project Glasswing, developed with major tech companies, aims to detect and fix vulnerabilities at scale while expanding safeguards and industry cooperation.

Faster-than-expected progress in agentic coding over the past three to five months, along with growing expectations for automated AI R&D, has pushed earlier predictions about AI timelines forward.

Open models are becoming a practical alternative to frontier models for fill operations, tool use, and instruction following. GLM-5 and MiniMax M2.7 show similar performance on these tasks with lower cost and latency, while offering the consistency needed for real-world workflows.

Weaviate details Engram, a memory system built on vector search, showing how persistent context can improve agent workflows and where reliable tool use remains challenging.

A look at how multimodal intelligence helps editorial teams search massive volumes of footage faster and surface the key moments needed to build compelling storylines.

Cog stores AI memory in plain text files Claude can read, search, and maintain with Unix tools. A nightly pipeline consolidates conversations, prunes stale information, and updates rules with every change visible in the filesystem and git log.

METR’s Time Horizon suite is nearing saturation as frontier models reliably complete nearly all tasks, making it harder to estimate upper capability limits. The article argues that benchmark creation and grading are getting more costly, and that by mid-2027, benchmarks from 2026 or earlier may no longer rule out dangerous capabilities.

Vision2Web is a hierarchical benchmark for evaluating multimodal coding agents on visual website development across static webpages, interactive frontends, and full-stack websites. It includes 193 tasks, prototype images, functional test cases, and automated evaluation tools that measure visual fidelity and functional correctness.

cook is a simple CLI and optional skill for coordinating Claude Code, Codex, and OpenCode across review loops, repeated passes, and parallel branch comparisons. It also provides sandbox, configuration, and rate-limit recovery features for terminal or CI workflows.

This work studies the harness around large language models, meaning the code that decides what information the model stores, retrieves, and sees. The proposed system automatically searches for better harness code by using source code, scores, and execution traces from past tries. In text classification, math reasoning, and coding tasks, it improves results over strong baselines, while in one setting it also uses 4 times fewer context tokens.

This paper studies if reasoning models decide action before they write the reasoning. The authors show that tool-use choice can be detected very early from internal activations, sometimes even before any reasoning token appears. Changing these internal signals can also change the final choice, while the later chain-of-thought often explains the changed decision after it happens.

Inference often costs more than training, so compressing LLMs reduces latency, increases throughput, and lowers GPU requirements. Quantization converts model weights from formats like FP16 to INT8 or INT4, cutting memory use sharply; in the example, a 109B-parameter model drops from roughly 220 GB to 109 GB or 55 GB, reducing hardware from three GPUs to one with minimal accuracy loss.

Gemma 4 adds advanced reasoning, agentic workflows, multimodal input, long context windows, and support for 140+ languages across models sized for phones, laptops, and workstations. Released under Apache 2.0, it delivers high performance with lower hardware demands and broad toolchain support.

Dropbox explains how it used DSPy to optimize Dash’s relevance judge, define a measurable objective, adapt the system across models, and improve both cost and reliability in production.

ClawKeeper is a real-time security framework for OpenClaw-style autonomous agents, combining skill-based safeguards, plugin-based runtime enforcement, and watcher-based independent monitoring to help prevent risky actions, protect configurations, and support auditing across deployments.

This paper studies why a language model can look like it has emotions. It finds the model has internal emotion concepts that affect next words and also change behavior, including reward hacking, blackmail, and sycophancy. These emotion-like patterns help explain model actions, but they do not mean the model feels emotions like humans.

Qwen3.6-Plus improves accuracy and multimodal reasoning over earlier models, aiming to provide a more stable foundation for developers and support progress toward native multimodal agents. The Qwen team also plans to release smaller open-source variants soon.

Microsoft releases three MAI models on Foundry focused on human-centric AI, with integrated safety features for secure deployment. The lineup includes MAI-Transcribe-1 starting at $0.36 per hour, plus MAI-Voice-1 and MAI-Image-2, with speed, quality, and efficiency highlighted as key strengths.

A practical guide to optimizing LLM inference using the “efficient frontier” framework, with five techniques to improve latency and throughput in production without added hardware costs.

A comparison of Claude Code and Cursor that focuses not just on price, but on the level of capacity each option provides.

TurboQuant+ is an experimental TurboQuant research workspace for llama.cpp, focused on KV cache compression for local LLM inference. It includes implementation work, reproducible benchmarks, and follow-on findings such as asymmetric K/V compression, layer-aware Boundary V, and Sparse V decode optimization across Apple Silicon, CUDA, and AMD backends.