Change-aware adversarial release testing

CI passes because nobody wrote the test. Gauntlet reads the change and writes it.

It plans attacks against each pull request, runs them on base and head, and shows you what the change broke.

How a Gate run decides
Gate pipeline: the PR diff goes to the planner, which proposes a plan that is hashed and frozen. The plan runs on base and head environments, an independent oracle compares them, and the verdict reports a regression.01PR diffbase..head02Plannerproposes attacks03Frozen plan9f2c…e41a04Base envisolated · ×204Head envisolated · ×205Oracleindependent check06Verdictregressionhead 2/2

Two products

Gate

Before merge: attack what the PR changed, on base and head.

3
Runs
2
Regressions
1
Passed

Recall

After it shipped: audit past agent decisions and replay them.

Live numbers unavailable while the API is offline.

How it stays safe

  • The model only proposes
    It never decides a verdict.
  • Plans are frozen
    Approvals bind to a hash.
  • Oracles decide
    Evidence, not the model.
  • Humans approve
    Nothing auto-merges.
  • Built with
  • Daytona
  • Anchor
  • Azure OpenAI
  • Arize
  • SurrealDB
  • ElevenLabs
  • GitHub
Sponsor proof