AI Security Is Starting To Look Like Social Engineering

When most people think about security, they imagine:

exploits

malware

vulnerabilities

unauthorized access

Traditional systems are usually attacked technically.

But AI systems are starting to behave differently.

The Strange Thing About AI Systems

While testing AI agents recently, one pattern kept showing up:

Many failures didn’t come from hacking.

They came from persuasion.

A small wording change.

A conflicting instruction.

A more convincing request.

And suddenly:

safeguards weakened

outputs changed

instructions were ignored

No exploit.

No malware.

No crash.

Just conversation.

AI Systems Respond To Language

That changes the security model completely.

Traditional software doesn’t “understand” persuasion.

AI systems do.

And that creates a weird new category of problems where:

tone matters

phrasing matters

instruction order matters

The system may technically function correctly—

while behavior still changes…