Execution-verified environments for coding agents

Proof, not preference.

Muence runs a real exploit against the code and the app's own tests. A model scores only when the attack dies — and nothing else breaks.

A real app runsAn exploit proves the bugTests verify the fix
muence / verify
VIBESEC-0784

$ muence verify patch

Evaluating auth/session-fix.ts

01Boot app and run specpass
02Replay saved exploitblocked
03Verify patched behaviorpass
04Check held-out testspass

Reward earned

all executable gates passed

+1.00

1,000

execution-verified tasks

4 gates

before a task can score

0 / 1

a program is the final judge

Why it matters

Code can work and still be dangerously wrong.

Training rewards functional output: does the app run? Security is usually absent from that loop. Models learn to ship a feature, even when the feature leaves an exploit behind.

Muence turns the missing signal into a runnable environment. The exploit must fail, and the original spec must still pass.

Inspect the benchmark

Where models fail

Failure rate across verified security tasks

Broken access control78% fail
Missing input validation71% fail
SQL / command injection65% fail
IDOR60% fail
A passing app is not enough. The exploit must be dead too.

The executable loop

A reward signal that has to survive contact with reality.

Every environment starts with a working but vulnerable app, then proves both the bug and the fix using isolated, executable gates.

01

Generate

Start with realistic product code: logins, uploads, payments, and APIs — written without security guidance.

02

Prove

A runnable exploit proves the vulnerability exists. A held-out suite defines the app behavior that cannot break.

03

Reward

The patch earns a score only when the exploit fails and the product still works. That outcome trains the model.

Built for measurement and training

One engine, two ways to use it.

Evaluate the frontier today, then use the same executable tasks as the training environments that move the score.

Evaluate

Execution-Verified Benchmark

A public security benchmark built from tasks with real exploits, held-out specs, and hard pass/fail results.

  • Boot · exploit · patch · re-verify
  • Open VibeSec dataset and live model rankings
  • Scores based on programs exiting 0 or 1
View the leaderboard
Train

RL Environments

The same tasks become sandboxed reset / step / reward loops, ready to generate trajectories for post-training.

  • Isolated execution with egress blocked
  • A functional grader blocks reward hacking
  • Rollouts export into post-training pipelines
Talk to us

Build with us

Make the reward worth optimizing.

Talk with us about the benchmark, data partnerships, or running your model through execution-verified environments.

MuenceMuence

Execution-verified environments for coding agents.