This Is Already Happening
These aren't hypothetical threats. Every finding below comes from public research published in the last 90 days.
19 npm packages with hidden backdoors targeting Cursor, Claude Code, and Windsurf
Malicious packages in the npm registry specifically targeted developers using AI coding assistants, injecting backdoors into agent workflows.
Socket Threat Research, Feb 202621 of 36 documented attacks exploit 4+ stage kill chains
The majority of real-world AI agent attacks chain multiple vulnerabilities together, making single-point defenses insufficient.
Schneier, "The Promptware Kill Chain," Jan 2026Only 14.4% of AI agents go live with full security approval
The vast majority of AI agent deployments bypass security review entirely, creating blind spots across identity, access control, and observability.
Gravitee API Security Report 202628 Controls. 3 Tiers. 6 Pages.
Each control maps to a real attack vector and a specific defense. Pick the tier that matches your deployment stage.
- ✓ Unique identity per agent instance Critical
- ✓ Secrets in vault, never in code Critical
- ✓ Input/output validation on every tool call High
+6 more controls
- ✓ MCP tool allow-list enforcement Critical
- ✓ Container/sandbox per agent execution High
- ✓ Rate limiting on LLM & tool calls High
+6 more controls
- ✓ Multi-agent trust boundary enforcement Critical
- ✓ Kill-switch for agent termination High
- ✓ Full audit trail with tamper-proof logging Medium
+7 more controls
Ready to secure your agents?
Get the full 28-control checklist as a free PDF.
No spam. No sales pitch. Just the checklist.