Snyk Labs

Imagine an agent that reads your email, pulls context from internal docs, and drafts replies. No memory-unsafe code. No SQL injection flaws. No suspicious endpoints. The permissions are valid. The tool calls are legitimate.

And it can still be compromised by a single sentence.

We are witnessing a fundamental shift in application security. Traditional tooling is built to find bugs in deterministic code. But generative AI systems do not fail at the level of code. They fail at the level of behavior.

In the deterministic world, we "solved" whole classes of security problems by enforcing hard boundaries. SQL injection became manageable because we could parameterise queries, separating "instructions" from "data." If your input remained data, the trust boundary held. With agentic AI, that trust boundary has dissolved.