Critical SQLite CVEs Turned Out to Be AI Fabrications

JFrog examined several vulnerability reports for SQLite and reached a disheartening conclusion: some of them simply don't exist. These CVEs were completely made up by large language models, yet they still managed to seep into authoritative vulnerability databases, which is where researchers came across them.

Garbage That Looks Like a Threat

The problem with such “hallucinations” isn't just that they clutter up the databases. Organizations spend real time investigating and trying to close holes that aren't there. This is especially painful in environments where critical vulnerabilities are automatically prioritized and tickets are opened right away based on the severity rating. On paper, a high rating for a nonexistent CVE turns into a genuine burden on the security team.

When an AI Agent Takes Over

A separate concern is scenarios where vulnerability handling is automated with AI. An agent that stumbles onto a fabricated CVE may earnestly try to locate the vulnerable function, generate a patch, or propose changes — except that it's relying on code that doesn't exist.

Instead of helping fix real problems, such an assistant leads specialists down a false trail. In the worst case, it introduces needless edits into working code and forces people to spend time untangling the consequences of its own activity. The result is a vicious circle: one neural network invents a vulnerability, another diligently “fixes” it, and living, breathing people are left to sort out the mess.

The SQLite story is a vivid illustration of how a stream of machine-generated “noise” can undermine trust in the tools that security teams have come to rely on. As long as verifying vulnerability reports remains the responsibility of humans rather than automated pipelines, incidents like this will keep cropping up more and more often.

Source: LWN.net