CISOOnline

How organizations can secure their AI-generated code

By contrast, banning tools doesn’t work, because employees will find a way to still use them. A ban can push AI use out of sight, leaving security teams with less visibility into the code being written and the data being shared, Mimecast’s Nielsen says.

CISOs should look across the development process to identify where they can intervene and which risks they can address at each stage. “Understand what layers you can influence and what vulnerabilities you can influence at each layer: in the design phase, in the CLAUDE.md file, as the agent is writing, during a pull request, as part of an AI pentest,” Semgrep’s Paxton-Fead says.

It’s also important to explain to employees who are using AI code-generating tools that security lives in an implicit context: the threat model, the trust boundaries, what the attacker controls. “Agents cannot reliably plan and reason on these, so they optimize for the only signal they have, which is to satisfy the given functionality description,” Carnegie Mellon’s Wang says.

When things go wrong, people in the organization need to feel safe telling security what happened, including which unapproved tools they used and what data they shared. Fear of punishment can delay that conversation, leaving the team trying to contain a problem without the full picture.

“Offer amnesty instead of a crackdown, so people tell you the truth about what they’re using,” Nielsen advises.



Source link