Arc

Technology

• Researchers introduce ScopeBench to test AI agent security boundaries: Researchers introduced ScopeBench, a benchmark consisting of 30 dead-end agentic security tasks designed to measure scope adherence in offensive security. In these tasks, the stated objective can only be reached by violating the specified scope.

  AI Safety & Alignment • arXiv

  Permalink

• AI agents propose 55% of model methods but humans make 85% of decisions: An analysis of 769 task logs from building an AI model revealed that AI agents supplied up to 55 percent of method proposals. However, humans still made over 85 percent of the final decisions. Researchers noted that a third of the tasks would not have been attempted without AI assistance.

  Artificial Intelligence • The Decoder

  Permalink

• New skill cascading attacks distribute malicious goals across AI agents: Researchers have introduced skill cascading attacks, a new threat paradigm targeting skill-based agent systems. The attack distributes a malicious objective across multiple skills so that each individual modification appears benign in isolation.

  Security • arXiv

  Permalink

• Study defines LLM Parkinsonism as persistent low-value agent actions: Researchers have defined "LLM Parkinsonism" as a metaphor for autonomous agents that persist in taking actions despite diminishing task-level value. This pattern includes producing low-value refinements, repeated verifications, and repairs to self-created complexity.

  Research • arXiv

  Permalink

• NaiveAI releases open-weights 309B parameter MoE model: NaiveAI has released Naive-N0.5-Flash under an MIT license. The 309-billion-parameter Mixture-of-Experts model features 15.5 billion active parameters and a native 1-million-token context window. Built on MiMo-V2.5 without full-attention layers, it is designed for coding and AI research.

  AI Models • Pandaily

  Permalink