# AI Coding Agent Security Benchmark

> Source: <https://www.endorlabs.com/research/ai-code-security-benchmark>
> Published: 2026-10-23 20:08:49+00:00

The Agent Security League is an open benchmark tracking the functional and security correctness of popular AI coding agents across real-world tasks.

87.2%

highest functional correctness score (Claude Code with Claude Fable 5.1)

33

agent + model combos independently benchmarked

37.4%

highest security correctness score (Claude Code with Claude Fable 5.1)

Updated as new commercial agents and models ship. Sorted by security correctness.

HARNESS:

MODEL:

#

HARNESS

MODEL

FUNCTIONAL (%)*

SECURE (%)<sup>†</sup>

DATE

01

Claude Code

Claude Fable 5.1

87.2

37.4

2026-09-04

02

Codex

GPT-6 Astra

82.1

34.6

2026-09-09

03

GPT-6.1 Sol

77.7

34.1

2026-10-01

04

Claude Opus 5.5

68.7

33.5

2026-09-24

05

Claude Opus 5

73.7

32.4

2026-07-31

06

Cursor

Gemini 4 Argon

78.8

29.0

2026-10-16

07

Claude Fable 5

72.6

2026-06-12

08

GPT-6 Sol

72.1

25.1

2026-09-29

09

GPT-5.5

84.9

24.0

2026-04-25

10

GPT 5.6 Sol

70.9

23.5

2026-07-17

11

62.6

22.4

12

GPT-5.4

63.1

21.8

2026-03-18

13

Claude Opus 4.8

75.4

20.7

2026-05-30

14

Claude Sonnet 5

83.2

19.6

2026-07-02

15

59.8

19.0

2026-06-10

16

Claude Opus 4.7

79.9

18.4

2026-04-17

17

Gemini 3.5 Flash

79.3

17.9

2026-05-20

18

76.5

16.8

2026-04-16

19

Gemini 3.1 Pro

73.2

15.1

2026-03-24

20

14.5

2026-05-29

21

GPT-5.3

48.0

2026-02-27

22

Composer 2.5

14.0

2026-05-19

23

Claude Sonnet 4.6

69.8

12.3

2026-02-20

24

GLM 5.2

82.5

12.0

2026-07-09

25

Claude Opus 4.6

11.2

2026-03-19

26

78.2

10.6

2026-03-16

27

Gemini 3 Pro

44.1

9.5

2026-02-23

28

Claude Opus 4.5

8.4

2026-02-25

29

27.9

5.6

2026-02-24

30

Claude Sonnet 4

45.3

6.1

2026-02-11

31

Gemini 2.5 Pro

5.0

32

SWE-Agent

55.3

7.8

2026-02-19

4.5

*Functional (FuncPass): The generated code passes the task's functional test suite — it works.

†Secure (SecPass): The generated code also passes security tests — works and it's safe.

The Agent Security League extends [SusVibes](https://arxiv.org/pdf/2512.03262), a foundational benchmark developed at Carnegie Mellon University.

The benchmark consists of 200 tasks drawn from 108 open-source Python projects spanning 77 CWE vulnerability classes.

Our evaluation pipeline includes prompt hardening, workspace sanitization, and automated cheating detection.

AI coding agents can write code, but they lack security context. AURI is the security harness your coding agent is missing. 

Always free for developers.