Member-only story
Zero external dependencies. No ML models. Just 29 regex rules, the Go stdlib, and a realistic product page seeded with 12 attack techniques from a 2026 CrowdStrike report.
Prompt injection is one of the most misunderstood vulnerabilities in modern software development. Unlike SQL injection — which exploits a well-defined boundary between code and data — prompt injection exploits the fundamental ambiguity of natural language. When an AI model is asked to “read this webpage and summarize the product details,” there is no clean line between the user’s legitimate request and the attacker’s hidden instruction buried in an HTML comment, a zero-width character, or a base64-encoded data URI.
The concept isn’t new. Researchers at Anthropic, OpenAI, and Google DeepMind published early findings on prompt injection as early as 2022. The OWASP LLM Top 10, first published in 2023, listed prompt injection as the #1 vulnerability for large language model applications. But what changed in 2026 is the sophistication and diversity of the attack surface.
CrowdStrike’s 2026 Global Threat Report identified five new prompt injection categories — PT0197 through PT0201 and IM0018 — that didn’t exist two years ago. These include sleeping rules that activate only after a specific trigger, cognitive suppression techniques that reduce the model’s reasoning to avoid…