Security/verification surge: Anthropic Glasswing + OpenClaw Marcus/YC risks + Kiro incdt + AgentHazard/ClawArena + Rubrik SAGE + ServiceNow + hallucinations + Kyndryl + OpenClaw safety + HDP + Ollama + testing
Key Questions
What is Anthropic's Project Glasswing?
Project Glasswing fixes agent vulnerability chains and OS escapes. It enhances security for AI agents.
What risks are associated with OpenClaw?
OpenClaw highlights risks debated by Marcus vs. Tan, including bans in Claude subscriptions. It raises safety concerns for AI agents.
What is the AgentHazard benchmark result?
AgentHazard benchmark shows computer-use agents fail safety tests at 73% rates. This underscores verification needs.
What is ClawArena?
ClawArena is a benchmark for AI agents in evolving information environments. It tests agent performance dynamically.
What is the reported hallucination rate?
Hallucination rates for frontier reasoning LLMs are around 4.6%. Claims of 'next to nil' are contested.
What is Rubrik SAGE?
Rubrik SAGE uses AI to govern and secure agentic workforces. It was featured at RSAC.
What ServiceNow vulnerability was reported?
ServiceNow AI platform had a 47% vulnerability in agent autonomy. It reflects broader enterprise cyber risks.
What does Kyndryl offer for AI agent security?
Kyndryl's Agentic Service Management provides a roadmap for AI agent automation and security, aligned with ISO standards.
Anthropic Glasswing fixes agent vuln chains/OS esc; OpenClaw risks (Marcus vs Tan); Kiro AWS downtime; AgentHazard 73%; ClawArena evolving evals; hallucinations 4.6%; Rubrik SAGE RSAC; ServiceNow 47%; DeepMind Traps 86%; Kyndryl ISO; HDP; IBM/Red Hat; Ollama honeypots; testing frameworks.