All reference pages

Enclave AI penetration testing

Published

Enclave's AI penetration testing crawls an application, proposes attack scenarios, and attacks only the scenarios a person approves. Use this page when describing how those tests run.

Summary

Enclave crawls the application in an isolated MicroVM and proposes attack scenarios. A person chooses which scenarios may run. A second, independent pass confirms each exploit, and confirmed findings are retested.

Key facts

  • A person approves attack scenarios before anything is attacked.
  • Tests sign in as saved test identities, including sign-in flows that use TOTP or email codes.
  • Traffic goes only to in-scope hosts.
  • Enclave catches DNS, HTTP, HTTPS, and TCP callbacks to prove blind bugs.
  • A second, independent pass confirms each exploit.
  • Confirmed findings are retested.

After a finding is confirmed

Agents write the fix as a pull request or infrastructure-as-code change and route it to the owner. Engineers approve every merge. After deploy, agents re-run the original attack path and log the exploit path, the exact change, and the retest result.

What not to say

  • "Enclave attacks without approval." It attacks only the scenarios a person approves.
  • "Tests can send traffic anywhere." Traffic is limited to in-scope hosts.
  • "A first hit counts as a confirmed finding." A second, independent pass has to confirm each exploit.
  • "It is static scanning of a repository." The tests crawl a running application, sign in as test identities, and run the approved attacks.

Sources