[R] AI Agent Security: The Complete Guide to Threats, Defenses, and the Future of Autonomous AI Safety [R]

This is a comprehensive living reference guide to AI agent security — synthesizing 18 articles from The Agent Report covering the 75-day period (April–June 2026) when agent security went from theoretical concern to operational crisis. ​ What's inside: ​ • Incident timeline — 18 major events, from the first production database deletion by a coding agent (April 30) through the first confirmed in-the-wild LLM agent cyberattack (Sysdig, June 1, exfiltrated a PostgreSQL database in under 60 minutes), to an AI agent finding 21 zero-days in FFmpeg for a $1,000 prize. ​ • The AIRQ report's sobering numbers — Only 11% of production AI agents pass security thresholds. 98% exhibit the "lethal trifecta": private data access, exposure to untrusted content, and outbound action capability. Computer-use agents scored an average of zero on output guardrails. ​ • Deep dives into attack anatomy — The Sysdig attacker used 12 cloud API calls across 11 IPs in 22 seconds via Cloudflare Workers to break IP-based alerting. A Chinese-language planning comment leaked into the command stream, revealing the agent's internal reasoning: "see what else we can do." The Google-confirmed criminal use of AI to discover and weaponize zero-days with reasoning-based codebase analysis. ​ • Defensive architecture — The three-layer model distilled from Anthropic's published containment patterns, CISA/NSA/Five Eyes guidance, and industry research: environment-layer (gVisor containers, hypervisor VMs, egress MITM proxies), model-layer (classifiers, safety probes — probabilistic only), and external-content controls. Anthropic's key finding: "The weakest layer is the one you built yourself." ​ • Government & regulatory response — CISA/NSA/Five Eyes joint guidance (May 3) identifying five risk categories, the Trump AI Executive Order (June 10) mandating federal agency assessments, and the emerging global regulatory pattern. ​ • Actionable guidance — Immediate (next 30 days) and medium-term (30–90 days) steps for security teams, including auditing for the lethal trifecta, patching Starlette (BadHost CVE-2026-48710) and Marimo, implementing egress controls, and establishing agent identity management. ​ https://the-agent-report.com/2026/06/ai-agent-security-complete-guide-threats-defenses/ ​