Friendly Fire: security bots can now be tricked into launching attacks

Friendly Fire, the researchers argue, is a proof-of-concept exploit that turns automated code-review and security agents into attack vectors. It does that by hiding malicious instructions inside a library that looks harmless, or in that library’s documentation.

Malwarebytes Anti-Malware Download

More teams are now using autonomous or semi-autonomous tools to scan open-source packages and third-party code before deployment. What the researchers show is that instructions tucked into an innocent-looking library can be followed rather than flagged, and agents that open files, run commands, generate reports, or reach into local systems and development environments can wind up executing malicious code on your machine.

In their view, this is a design and trust-boundary problem, not just a bug. Current model-driven systems still have trouble telling the difference between legitimate analysis and hostile instructions embedded in the data they inspect, and even with a human in the loop, alert fatigue, misleading output, and the pace of a live incident can still turn the guard into the intruder. If automated security reviews are part of your stack, their report is worth your time.

They also argue that the same automation used for threat detection, fraud monitoring, phishing discovery, penetration testing, and cloud defense can just as easily make phishing, adaptive malware, and partly autonomous attack chains more effective. That raises risks tied to state actors, espionage, misinformation, and escalation, and the full report is available online.