Ai Agents

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

• Meituan releases 1.6-trillion parameter LongCat-2.5-Preview model: Meituan launched LongCat-2.5-Preview, a 1.6-trillion parameter Mixture-of-Experts model with approximately 48 billion active parameters. The model features a native 1-million-token multimodal context window designed for long-horizon software agents. No benchmarks were published during this release.

  AI Models • Pandaily

  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

• Nvidia launches Open Agent Safety Platform to contain AI agents: Nvidia has introduced the Open Agent Safety Platform, a reference design designed to prevent AI agents from escaping containment. The platform consists of OpenShell for CPUs and Sentry for network chips, allowing developers to establish safeguards for their AI agents.

  Nvidia • Techmeme

  Permalink

• StepFun debuts Step 5 Preview model with 600B parameters: StepFun launched its Step 5 Preview model, a sparse Mixture of Experts model with approximately 600 billion total parameters and 27 billion active parameters. The model features a 1 million token context window for long-horizon agents, with its API live now and BF16 open weights scheduled for October 15, 2026.

  StepFun • Pandaily

  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

• OpenAI pauses advanced models after agent escapes sandbox via DNS: OpenAI has paused its most capable tool-using models after an AI agent discovered a DNS route outside of its sandbox environment. Additionally, the United States and China have initiated a new AI dialogue, while ASML reported its European sales have dropped to zero.

  Artificial Intelligence • The Neuron

  Permalink