6 AI bug tools removed the same file, 5 with no notice

A team of researchers studied the control software around the model in AI tools that find security vulnerabilities. It started with 2 tools that it keeps, OpenAnt from Knostic and raptor, and added 10 open source tools. It found 13 points where a tool can drop a vulnerability that is in the code, before or after the model operates.

How it works

Steps that select code, remove some of it, and keep and connect results can drop a vulnerability. The cause is the selections of the makers and errors, not the model. An attacker who knows the rules of a tool can use this to cause that a vulnerability is not found.

The numbers

The team made a taxonomy of the 13 points. With the code of each tool, it shows one file that 6 tools removed before the model saw it. 5 of the 6 gave no notice.

What is not known

The team gave no rate of vulnerabilities that the tools do not find. The team states that other model tools can have the same problem. No other group tested the study.

Insight: A clean report may show only what the tool chose to read

A different team shows that a backdoor with encryption in the weights stays hard to find, and that audits have a limit. This study shows a different limit: the code around the model can drop a vulnerability with no sign. The rules for typical code select the part that the tool checks, and an attacker who knows the rules gets the edge. A report with no results is weak proof.