Cerberus
Cursor for AI hacking that can't go out of scope
62 followers
Cursor for AI hacking that can't go out of scope
62 followers
Cerberus is the world's first safe AI hacker. You can hack your entire app in plain English with a prompt "find vulnerabilities and exploit them in example.com". We also built the world's first AI hacker that's mathematically safe to run on production. It uses a new programming language where every hacking action must come with a mathematical proof that you authorized it — no proof, no action. Point it at your app, come back in 3-4 hours with a full security report.
Cerberus
@aziz_akhmedkhodjaev This will have huge application in Third Party Risk Space. a lot of small vendors cut corners when it comes to pen test. this would definitely help them becoming compliant
Ichiba AI
Scope question: does this cover prompt-injection and context-manipulation attacks, or is it authorization-boundary focused? Those are very different security problems.
Cerberus
@ichiba Hi, thanks for the question.
It covers all kinds of attacks that may lead the AI hacker to do unsanctioned (out of scope) actions, because we mathematically model the scope of penetration testing.
For example, you might have allowed only testing of *.yourdomain.com subdomains and disallowed DDoS attacks. Any program code that will attempt to launch an attack to something else outside of *.yourdomain.com or would start a DDoS attack won't pass type checking and will be caught off during compilation.
jared.so
Formal-methods-enforced scope ("no proof, no action") is the only sane answer to prompt-injected AI hackers going rogue. A purpose-built typed language where out-of-scope code doesn't compile is the first AI pentest tool that could actually run against prod without a lawyer on speed dial. Curious how the language handles emergent attack chains where each step is individually in-scope.