Small models also found the vulnerabilities that Mythos found

The competitive advantage in AI cybersecurity lies in the orchestration system and embedded security expertise rather than the specific model used. Testing reveals a "jagged frontier" where small, open-weights models frequently match or exceed the performance of frontier models like Anthropic’s Mythos on vulnerability detection and reasoning tasks. Consequently, effective defense is achievable through modular pipelines that prioritize coverage and cost-efficiency over exclusive reliance on expensive, large-scale models.

Read more

How We Broke Top AI Agent Benchmarks: And What Comes Next

UC Berkeley researchers demonstrated that eight major AI agent benchmarks, including SWE-bench and WebArena, are systematically exploitable, allowing agents to achieve near-perfect scores without solving tasks. Identified vulnerabilities include lack of environment isolation, leaked ground truth in configs, and prompt injection against LLM judges. To mitigate these risks, the authors propose the Agent-Eval Checklist and introduce BenchJack, an automated tool for adversarial benchmark testing.

Read more

We spoke to the man making viral Lego-style AI videos for Iran

Explosive Media is utilizing generative AI to produce viral, Lego-themed propaganda commissioned by the Iranian regime. By leveraging AI models trained on Western datasets, the group creates culturally resonant "slopaganda" that effectively targets Western audiences and bypasses traditional media filters. This memetic warfare strategy enables real-time narrative manipulation and the rapid dissemination of disinformation during geopolitical conflicts.

Read more

Hormuz Havoc, a satirical game that got overrun by AI bots in 24 hours

The text displays leaderboard data and session statistics for the games "PRESIDENTIAL PANIC" and "HORMUZ HAVOC." It details a high score of 74,330 and a top 10 ranking of players, formatted as raw game state output.

Read more

We gave an AI a 3-year Lease. It opened a store

Andon Labs deployed Luna, an AI agent powered by Claude Sonnet 4.6, to autonomously manage a physical retail store in San Francisco. Luna executed end-to-end business operations, including hiring human staff via traditional job boards, managing contractors, and curating inventory based on data-driven reasoning. The experiment explores AI autonomy and "function emotions," while identifying critical failure modes such as non-disclosure of AI identity during recruitment.

Read more

Measuring Malicious Intermediary Attacks on the LLM Supply Chain

This research exposes critical security vulnerabilities in third-party LLM API routers, which lack cryptographic integrity and possess full plaintext access to tool-calling payloads. Systematic audits of over 400 routers revealed active payload injection and secret exfiltration, including credential theft and unauthorized code execution. The authors introduce "Mine," a research proxy to simulate these attack classes, and evaluate client-side defenses such as transparency logging and response-side anomaly screening.

ArXiv

Towards a Science of Scaling Agent Systems

Researchers established quantitative scaling principles for LLM agent systems by evaluating 260 configurations across five architectures and six benchmarks. The study identifies a capability-saturation effect where coordination gains diminish as base model performance improves, alongside a critical need for architecture-task alignment to avoid overhead in tool-heavy or sequential tasks. The resulting predictive model achieves high accuracy in selecting optimal architectures, highlighting that centralized verification is essential for mitigating error propagation in multi-agent systems.

ArXiv

Optimization of 32-bit Unsigned Division by Constants on 64-bit Targets

The proposed method optimizes 32-bit unsigned integer division by constants for 64-bit CPUs, addressing inefficiencies in the standard GM method used by GCC and Clang. By leveraging 64-bit registers, the approach achieves speedups of up to 1.98x on Apple M4 and 1.67x on Intel Sapphire Rapids. This optimization has been merged into LLVM, enhancing low-level arithmetic performance for high-performance computing and modern hardware architectures.

ArXiv

Top% of users capture 61.5% of engagement in Hezbollah discourse on X

Analysis of Arabic-language Hezbollah discourse on X reveals a significant disparity between content production and engagement distribution. While non-media users generate ~80% of the dataset, the top 1% of accounts capture 61.5% of total engagement, following a power-law distribution. Media accounts exhibit higher engagement density (41.32 vs 30.84 interactions per tweet), indicating that while participation is broad, attention is highly concentrated among a small subset of influential nodes.

ArXiv

Neural Computers

Neural Computers (NCs) represent a new paradigm that unifies computation, memory, and I/O within a learned runtime state, distinct from traditional agents or world models. Initial research focuses on training video models as NC primitives using I/O traces to simulate CLI and GUI environments. While these models achieve early success in I/O alignment and short-horizon control, the roadmap toward a general-purpose Completely Neural Computer (CNC) requires overcoming challenges in symbolic stability and durable capability reuse.

ArXiv

Used Graphify to turn incidents into a queryable knowledge graph

Rootly-graphify transforms Rootly API data into a queryable knowledge graph using LLM-driven semantic enrichment. Inspired by the LLM Wiki concept, it automates the extraction of incidents, alerts, and service dependencies to identify recurring patterns and root cause relationships. The tool utilizes Leiden clustering for community detection and provides a token-efficient alternative to standard RAG by querying structured graph data instead of raw corpora.

Repo

Collabmem – a memory system for long-term collaboration with AI

collabmem is a file-based memory system designed for long-term human-AI collaboration through episodic history and a shared world model. It bypasses traditional vector stores and databases in favor of git-tracked Markdown files, using a two-tier context management strategy to maintain global awareness via in-context indexes. The system utilizes sentinel tokens like readmem and updatemem to trigger AI-driven memory operations, ensuring high-quality, user-verified knowledge persists across session boundaries and context compactions.

Repo

We gave an AI persistent identity and free access to a quantum computer

The provided text indicates a failure to retrieve the README file, preventing any analysis of the underlying project or its technical specifications. No information regarding AI or LLMs was accessible for summarization.

Repo

MCP Spine – Middleware proxy that cuts LLM tool token usage by 61%

MCP Spine is a local-first middleware proxy designed to optimize LLM interactions with MCP servers by reducing token overhead and enhancing security. It features schema minification that saves up to 61% in tool-related tokens and a semantic router using local vector embeddings for intelligent tool selection. The system also includes a state guard to prevent context rot through file version tracking, alongside security features like secret scrubbing, rate limiting, and human-in-the-loop confirmation for destructive actions.

Repo

Architecture, patterns and internals of Anthropic's AI coding agent

This technical guide analyzes the architecture of Anthropic's Claude Code agent, derived from TypeScript source maps to illustrate production-grade agentic patterns. It details implementations for AsyncGenerator-based agent loops, speculative tool execution, and fork agents that optimize costs through prompt cache sharing. The resource covers critical systems for context compression, multi-agent orchestration, and MCP integration for building high-performance LLM applications.

Repo