If AI "uncovered" it, that means it was already in the training data and likely exploited in the wild.

So with that logic, AI won't find a vulnerability in my newly written code, since that particular vulnerability wasn't in its training data?

That's... not how coding agents work.

cowlevel 2 hours ago | parent | next [-]

Sometimes it is. If the AI can do it for the defender it can also do it for the attacker.

	▲	trollbridge an hour ago \| parent [-]
		I wish. As the defender, I have a model that just barks at me that I hit a guardrail. The attacker has no such constraints.

varispeed 2 hours ago | parent | prev [-]

They guess the next token based on probability which is derived from training data.