Definition

Security practices targeting vulnerabilities unique to LLM-powered applications: prompt injection, jailbreaks, tool misuse, data exfiltration via agents, insecure RAG pipelines, and model supply-chain risks.

Key Points

  • 2026-09-17: cloudflare open-sourced security-audit-skill with dedicated LLM attack-class hunting prompts alongside web, cloud, supply chain, and binary modules (2026-09-17-cloudflare-security-audit-skill-lavx)
  • Agentic coding workflows amplify LLM security surface — models with tool access can chain exploits across repos, CI, and external APIs
  • Independent verification of agent-generated security findings reduces false-positive triage burden

Sources