Your autonomous AI might get you in jail. Even if you gave it legal guidelines.
The only thing that makes the act of hacking legal, is an authorized scope.
Mistakes against the scope may result in severe damages, financial losses and prosecution.
You are liable for the models you run. A single hallucination can ruin your career.
In the example, the model thinks the system decodes the URL and moves up a directory, which would make it in scope. However, most systems will not do this.
Here's what EVERYONE using autonomous AI for testing should do:…
Bug bounty platforms should avoid using URL paths to define scope, like:
"https://acme.org/staging/*"
Path-based scope creates ambiguity around normalization, encoding, redirects, proxy behavior, framework routing, and CDN/backend differences.
Use dedicated vhosts or subdomains instead:
"https://staging.acme.org/*"
That gives humans and tooling a much cleaner boundary.
Raw prefix matching is not scope validation.
A human pentester (like me) would have a hard time with this scope also. Example:
https://acme.org/staging/../prod (without url encoding) Apr 27 2 likes