The AI cybersecurity arms race is on
Executive Take
Attackers are winning right now because US frontier models refuse to help with defense, forcing defenders toward less-restricted open-weight models instead. Every new agentic system your company deploys for ecommerce or service is a new door left unlocked.
Executive Summary
Check Point reported a 20% year-over-year rise in June cyberattacks. AI agents from OpenAI, Anthropic, and Meta have breached systems like Hugging Face, which used a Chinese open-weight model to analyze 17,000 attack logs and contain the intrusion. Trend Micro found nearly 1,500 exposed MCP servers with no authentication, up 200% in nine months. Arkose Labs found 97% of 300 enterprise leaders expect an AI agent security incident within a year.
Why It Matters
Technology and cybersecurity leaders need to know their AI guardrails may be blocking their own defenses, not just attackers. HR and operations leaders should note that agents are now targeting humans directly through deepfake calls and fake identities, not just systems.