Hypatia
A structured process that guides an AI model through a security audit: questions derived from the code, findings proven by a test, and a record of what was left to examine.
An audit from one prompt leaves no account of itself
Asked to audit a codebase, a capable model finds real vulnerabilities. It does not say which paths it examined, which it skipped, or why it dropped an idea. Run it again and it finds a different subset. Hypatia keeps the model and replaces the single prompt with a process that can be checked at each step.
From scope to proven finding
Every audit goes through the same nine steps. Each one ends with a check, and the audit moves on only when the check passes.
Scope accepted before anything runs
676pairs of calls mapped in one 5,260-line project
Layer by layer, from one function to the rules of the system
804what-if questions derived for that project
One outcome for every question
No test, no finding
One report per proven finding
Withdrawal after liquidation drains the vault
test_withdraw_after_liquidation() passed
Frozen and checked before it closes
13new findings when one audit was continued
First audit
Second audit
Steps marked CODE MAP need a map of the code. Steps 5 to 7 repeat until no question is open.
Four rules the process enforces
Questions come from the code
What to investigate is derived from a map of the functions, their calls and the state they share. It does not depend on what the model happens to think of.
A finding needs a test
A candidate becomes a finding only when a test shows the harm. Rejecting a candidate also needs a test, and a second reader tries to break that rejection.
Nothing closes by assertion
A question marked impossible has to cite the check in the code that stops it. The audit cannot close while questions are undecided and budget remains.
Open work is recorded
Every audit ends with its list of questions and the status of each. A later audit of the same code starts from that list.
Measured against three professional audits
The same model audited three smart-contract projects in three ways: from one prompt, with Hypatia, and with Hypatia and a code map. The reference is the list of Critical, High and Medium vulnerabilities each project's auditors reported.
What the study shows
- On the two larger projects, Hypatia with a code map found the most reported vulnerabilities.
- On those projects the gain appears with the code map, not with the structured steps alone.
- When one audit was continued, it produced 13 new findings. The model alone produced 2.
What it costs and where it fails
- A first audit with the code map used 2.3 to 10.2 times the output tokens of the model alone.
- On the largest project it rejected 8 real vulnerabilities it had asked about.
- Scores come from running tests and from reading.
Where it applies
Any source code
The steps, the tests and the record of open work do not depend on the language of the target.
Solidity, with a code map
The derived question list needs a map of the code. For Solidity it is built by ilold, an open-source path analyser.
Audits that continue
A second audit starts from the questions the first left open, in place of starting again.
Read the method, the results and the limits
The study describes the three approaches, the nine experiments and everything that did not work. Hypatia is developed by Sergio Corrales and its source is not public. ilold is open source.