Search Results for “proof of work” – Page 17

Local Monitors Are Provably Blind to Multi-Agent Attacks — and the Standard Safety Net Has a Mathematical Hole

Enterprise security teams currently deploying multi-agent systems are operating under a false sense of security. As organizations increasingly integrate frameworks like CrewAI, AutoGen, LangGraph, and OpenClaw, they rely on per-step runtime monitors to catch malicious activity. New research, When Local Monitors Miss Compositional Harm: Diagnosing Distributed Backdoors in Multi-Agent Systems (Hu and Wang, 2026), reveals…

Meta Enters the Paid API Market With Muse Spark 1.1 at $1.25/$4.25

Meta’s first paid developer API model undercuts Claude and GPT by roughly 4x. Independent benchmarks confirm agentic tool-use strength—but Meta’s own reported scores are drawing skepticism.

Zhipu Doubles Down on Open Access as Beijing Considers Closing the Door

GLM-5.2, released under MIT license with a 1M-token context window, is the strongest open-weights model available. The timing—days after a U.S. government directive restricted Anthropic models—is deliberate positioning.

GhostApproval: Symlink Flaw Tricks 6 AI Coding Assistants Into Writing to SSH Keys — Anthropic Disputes, Then Ships Its Own Fix

Wiz Research discloses a category-level vulnerability in AI coding assistants where malicious symlinks bypass approval dialogs to write to sensitive system files. Five of six vendors fixed or acknowledged. Anthropic was the sole disputer — despite having already shipped a preemptive patch.