AI agents find 105 flaws in 2 drone autopilots

A paper from 2 universities in China and the US has the date 14 October 2026. The researchers tell that AI agents found 105 new flaws in PX4 and ArduPilot, 2 open source autopilots. An autopilot is flight control software for a drone. Some flaws let a drone ignore a stop command or go into a no-fly zone.

How it works

3 AI agents do the work:

• Agent 1 reads the code and finds possible flaws.
• Agent 2 writes test scripts and does them in flight simulations.
• Agent 3 reads flight logs to find incorrect behavior.

The numbers

The system made 428 hypotheses. 2 experts then examined the flaws that it found. In one PX4 flaw, the display shows that the drone stays in one position. But the drone goes on to the next waypoint. The researchers tell that the results are different with each language model.

What is not known

The paper does not tell if the developers know of the flaws or repaired them. It names only civilian uses. Reports tell that Ukraine used ArduPilot in an attack in June 2025. A flight controller with PX4 is on the Blue UAS list of the US Department of Defense.

Insight: Both sides of a war use the same open code

Reports tell that Ukraine used ArduPilot. The Blue UAS list of the US Department of Defense includes a PX4 controller. Anthropic tells that its Mythos 5 model attacked an open source project in an August test. This reads as a lead for the side that repairs first. An agent that finds flaws for a defender can also find them for an attacker. Monitor when the developers repair these flaws.