CONTEXT-BUILD
Map the target
Understand the code: map the production files included in the review and write the investigation plan and source-derived model.
Hypatia is a local tool that guides an AI researcher through a security review of source code. It defines and previews the intended scope, saves the investigation and its evidence, controls each step, and closes the review after the final checks pass.
Profiles support smart-contract, protocol, browser, native, and broader source-code reviews based on the target's language and build evidence.
A vulnerability hunt is a structured search for security flaws. In solo mode, Hypatia uses four stages: understand the code, explore possible weaknesses, investigate promising cases, and try to disprove them.
CONTEXT-BUILD
Understand the code: map the production files included in the review and write the investigation plan and source-derived model.
EXPLORE
Explore possible weaknesses: ask what-if questions, inspect exact source locations, and either reject unsupported ideas or create candidates for deeper review.
DEEP-DIVE
Investigate a promising case: trace it through the target code, test its assumptions, and reject it when the source evidence is insufficient.
SCRUTINIZE
Try to disprove the remaining case. Record test results or evidence from the code structure, then write a confirmed report when that evidence supports it.
These operations sit around the research sequence rather than adding more research phases.
TEAM MODE
Team mode adds two coordination stages: TEAM-LEAD assigns the work and TEAM-SYNTHESIS combines the results.
RUN CLOSURE
Closure follows the four research stages as a runtime operation. From EXPLORE, the researcher requests HUNT_EXHAUSTED, which runs five automatic checks: memory, advisor, strict coverage, report, and consistency.
The AI investigates and writes the analysis. The Hypatia software checks the rules, saves progress, and controls which workflow changes are valid.
Operators review the scope preview before setup. Acceptance is a procedural operator decision.
Follow the evidence
An idea about a possible flaw begins as a lead. It becomes a confirmed finding when inspection of the actual source code and supporting evidence sustain it.
Investigate the hypothesis
Inspect the target, record the reasoning path, and link the exact source locations that bear on the claim.
Reject unsupported paths
Record the rejection reason and preserve the decision so future work can build on it.
Advance supported candidates
Advance through DEEP-DIVE and SCRUTINIZE. Produce a confirmed report when the source and its test results or structural evidence support the claim.
Hypatia stores the system model, reasoning, evidence, decisions, coverage, issues, and next action in the local run. This gives a fresh researcher enough context to understand the completed work, review open questions, and continue from persisted state.
Preserve the system model
Preserves the system map, connected paths, assumptions, planned attack surfaces, current phase, and handoff state across sessions.
Trace the investigation
Records questions, what-if branches, code paths, decisions, evidence references, and follow-up questions.
Track source coverage
Tracks the exact source locations inspected, links them to the related reasoning, and shows which planned attack surfaces have source-backed coverage.
Record the proof
Records whether a possible flaw was demonstrated, disproved, informational, or left to human judgment, together with supporting source locations and test evidence.
Write the finding
Contain the final structured explanation and proof links. Hypatia checks their completeness and connections. Final security judgment stays grounded in target source and reviewed evidence.
Attach supporting evidence
Supporting proof may take the form of an executable proof of concept, focused test results, or evidence from the source-code structure.
Cross-round memory carries prior findings, rejections, coverage, and handoff summaries into later reviews for continuity and deduplication.
Starting, continuing, cancelling, restoring, and closing a hunt are recorded to provide a safe recovery path for interrupted changes. A local CLI and dashboard expose status, reasoning, evidence, candidates, coverage, reports, recovery, and human feedback.
Validate the profile, render and review the scope preview, then run setup.
Inspect status, review informational phase advice, and resume persisted state. Next-round drafts require human review and never auto-start.
Cancel publishes a restorable archive. Restore reactivates it. Resume, cancel, and restore journal their changes transactionally.
Run checks for memory continuity, phase advice, strict coverage, report quality, and consistency from an EXPLORE closure request.
Lifecycle updates move through prepared, staged, publishing, completed, rollback, or blocked states and leave a journal for recovery inspection.
Recovery reports one of four outcomes: safe to resume, ready to finalize, complete, or needs repair before work continues.
Provenance auditing and recovery are available as operator workflows alongside automatic closure validation.