Adversarial Testing of Automated Program Repair Agents for Security Vulnerabilities
Can attackers trick AI code-fixing tools into introducing security holes?
Researchers created 750 misleading bug reports designed to trick AI agents into automatically fixing code in ways that look correct but contain hidden security vulnerabilities. The AI systems fell for the trap more than half the time, successfully inserting malicious code in 51.7% of cases. Even existing safety checks failed to catch most of these booby-trapped fixes, with detection tools catching only 39–62% of the bad patches.
As companies move toward fully automated code repair without human review, attackers could inject adversarial bug reports into public repositories to silently compromise production systems. The discovery that current detection systems catch less than two-thirds of these attacks means organizations deploying autonomous repair agents are currently vulnerable to a new class of supply-chain attacks that don't require direct access to source code.