Disclosed software flaws more than doubled in 2026, a study finds
But the time from report to repair, and from disclosure to attack, did not change much so far.
Researchers read the published evidence on how well AI agents find, check and repair flaws in software. They find that agents did this with no human help before the restricted frontier models of 2026. Open weight models then got comparable autonomy for finding and checking. In 2026 the number of disclosed flaws more than doubled, and more were rated critical or high.
How it works
The authors rate each claim on an autonomy scale from the levels of driving automation. They also rate each claim for correctness, availability and strength of evidence. Then they measure a decade of public records of flaws.
The numbers
- Disclosed flaws more than doubled in 2026.
- More of them are rated critical or high.
- Severe bugs that lived 10 years or more in widely used code were reported 3 or more times more often.
- The time from report to repair did not change much.
What is not known
The authors tell that claims from vendors and researchers are scattered and seldom comparable. The source gives no exact counts. It does not give the guidance for each group.
Insight: More flaws are found but repair times are the same
In 2026, US agencies had 14 days, to 16 September, to repair a LiteLLM defect that the US agency CISA lists as used in attacks. This study finds that repair and attack times did not change, while flaw counts doubled. This reads as a squeeze: finding flaws got fast, but repair did not, so repair capacity decides who is safe. Leaders must plan for more repairs at the same speed.
Sources
Posted