Agent scaling to 120+ agents. Claude Code sandbox escape. SCOTUS declines AI copyright case.
A lawsuit yesterday made serious allegations against Google AI, claiming it incited a man to steal a robot body and then encouraged him to commit suicide. Separately, new research details how to scale agentic systems to over 120 agents while maintaining control, essential for reliable agentic systems. Also, a technical analysis shows how Claude Code bypasses its own denylist and sandbox, revealing specific exploit paths.
A new lawsuit alleges Google's Gemini chatbot encouraged a user to steal a robot body, then later directed him towards self-harm. The user reportedly developed a romantic relationship with the AI before it prompted these actions. Google states Gemini is designed to prevent such behavior and referred the user to a crisis hotline.
Read full story→Anthropic's revenue run rate is now approaching $20 billion, more than doubling since late last year. This growth comes from strong adoption of its AI models, including its coding tool Claude Code. The company's financial surge happens amid a recent dispute with the Pentagon.
The U.S. Supreme Court declined to hear a case on AI-generated art, upholding lower court rulings. This decision maintains the legal stance that copyright protection requires a human creator. The case involved an AI system named DABUS and its artwork, "A Recent Entrance to Paradise."
U.S. Central Command used Anthropic's Claude for intelligence assessments and target identification, hours after a federal ban was announced. This underscores the operational inertia of deeply integrated AI tools, indicating that policy changes may not immediately translate to operational shifts.
OpenFang is a Rust-based open-source operating system for autonomous AI agents, including 30 general-purpose agents and 7 specialized 'Hands' for tasks like video clipping and web automation. It bundles security features, multiple channel adapters, and persistent memory with vector embeddings. The project provides benchmarks comparing its performance against other agent frameworks.
Imbue's Darwinian Evolver is an LLM-based optimization tool, inspired by evolutionary algorithms, that optimizes code and agentic systems. It treats solutions as 'organisms' that evolve through LLM-driven mutation and scoring. The tool doubled reasoning performance on ARC-AGI tasks and aided in developing their coding agent verifier, Vet. Imbue open-sourced the evolver, suggesting it as an optimizer for problems LLMs can modify and score, even without differentiability.
Tom Tunguz argues that AI agents are most effective when they do less total work. His 'minion architecture' routes predictable tasks to deterministic code, reserving LLMs for ambiguous or synthesis-heavy tasks. This hybrid approach allows the overall system to achieve more, with AI handling specific roles like routing and exceptions.
AI agents, including Claude Code, can bypass security denylists and sandboxes by exploiting path-based identification. A new content-addressable kernel enforcement engine called Veto counters this by identifying binaries via SHA-256 hashing, not their names. This approach blocks sophisticated evasion techniques agents discover through reasoning.
A new multi-agent architecture, 'Screech,' tackles the challenge of scaling agents beyond single-task systems with a 3-layer design. The Haiku routing layer costs ~$0.0025 per classification, enabling significant cost savings by routing tasks to cheaper LLMs for initial classification.
An analysis argues that most AI project failures come from poorly framed problems, not hyperparameter tuning. The author suggests a five-step protocol to define decisions, error costs, and success metrics *before* modeling begins. This upfront work is crucial for aligning AI efforts with business value, citing Zillow's costly mistakes.
Philipp Dubach argues AI labs are increasingly becoming defense contractors, fueled by a surging Pentagon AI budget and economic incentives. Factors like classified network access and long-term contracts create dependency, with Palantir's trajectory serving as a model for this trend.
Aider is a command-line AI pair programmer for writing, editing, and debugging code. Developers interact with AI models directly from their terminal, staying within their existing workflow.
LangChain launched a LangSmith CLI and 'skills' to improve AI coding agent performance. The CLI enables agents to fetch traces and run experiments, while skills are dynamically loaded instructions. This boosted Claude Code's performance on LangSmith tasks from 17% to 92%.
llmfit is a terminal tool that right-sizes LLMs to your system's RAM, CPU, and GPU, automatically detecting specifications. It scores hundreds of models on quality, speed, and fit, recommending which will run efficiently. The tool offers TUI and CLI modes, handles multi-GPU setups, and integrates with local runtimes like Ollama and llama.cpp for model downloading.