June 20, 2026 · Jane Okafor, Marco Reyes

Securing AI agents before they secure themselves

Agentic systems fail in new ways: prompt injection, tool misuse, data exfiltration. A practical checklist for shipping agents safely.

An agent with tool access is an intern with root. Treat every capability grant as an attack surface.

Threat model at a glance

Threat	Vector	Severity	Mitigation
Prompt injection	Untrusted content in context	Critical	Boundary markers + output policy checks
Tool misuse	Over-broad capability grants	High	Least-privilege allowlists, human-in-the-loop
Data exfiltration	Outbound tool calls	High	Egress filtering, DLP on tool responses
Memory poisoning	Persisted adversarial state	Medium	Provenance tags, TTL on learned facts

Scope tokens, sandbox execution, log every tool call, and red-team the prompt boundary before launch.

policy.ts ts

export const policy = {
  allowTools: ["search", "read"],
  denyTools: ["shell", "email"],
};