Your protocol will beattacked. The onlyquestion is by whom.
Kerbex is an autonomous auditor for EVM protocols, built around evidence rather than opinions. 50–100 specialist agents hunt in parallel. A runner they cannot touch compiles and executes real exploit code against your repository. Roughly 150 deterministic control programs sit between the reasoning and your report.
- Accounting drift on partial withdrawproven
- Invariant holds on every path but oneproven
- Same defect repeated in three contractsproven
- Oracle staleness window unboundedreasoned
Confirmed findings
$1.32B was stolen from crypto protocols in the first half of 2026 alone, across 344 disclosed incidents — the most-hacked half-year on record. More than one protocol breached every single day.
Every one of those teams believed their code was fine. Most had been audited.
Don't let your protocol become a line in that table.
| Protocol | Root cause | Loss |
|---|---|---|
| BalancerNov 2025 | Rounding direction flaw | −$120M |
| Makina FinanceJan 2026 | Oracle manipulation via flash loan | −1,299 ETH |
| Verus BridgeMay 2026 | Validation flaw | −1,625 ETH |
| Solv ProtocolMar 2026 | Reentrancy — unaudited, no bounty | −$2.7M |
| Your protocolunaudited since | — | — |
How a finding gets from your commit to your report.
Seven stages, each producing an artifact the next one is allowed to depend on. Nothing advances on a model's say-so.
Book an auditGround truth first
A map of your protocol built from the compiler's own output — declarations, call paths, state reads and writes, value flows, entitlements, lifecycle transitions — each bound to an exact source span. Your real tests, fixtures and deployment scripts are mined too, so later proof work starts from how the protocol is actually used.
Parallel discovery
Several kinds of analyst run at once because each is blind to what the others see. First-principles lanes run in enforced isolation — they never receive pattern results or static output, so they cannot be anchored by them.
Convergence, then depth
Findings are fused and de-duplicated with full lineage, then checked for the same defect elsewhere — one reentrancy repeated in three more contracts is four findings, not one. Depth agents iterate up to three times, and a blind-spot pass reads the coverage matrix and attacks whatever nobody looked at.
Proof owned by the machine
Agents author exploit tests. They cannot run them. A supervisor executes every attempt inside a disposable clone, captures the result, tears the clone down, and verifies your source is untouched. A pass with no receipt is not evidence.
One accountable verifier
Each finding is adjudicated by a single fresh reviewer with clean context and direct repository access — no upstream verdict, no vote, no confidence score, and barred from reading its peers. Anchoring is the failure mode being engineered out.
Deterministic publication gates
A trust-model test, an impact-feasibility certificate, verbatim severity-matrix quoting from a protected snapshot, and citation resolution. Critical and High claims need executed, machine-attested proof to enter the headline set.
Nothing silently disappears
Every candidate gets a stable identity at birth and is tracked to a terminal state. The run is governed by an append-only, hash-chained journal that is the single source of authority — and it cannot be certified clean while one obligation is open.
Depth you can count.
Specialist agents per engagement, across non-overlapping discovery lanes
Deterministic control programs between the AI and your report
Average autonomous discovery and exploit-construction phase
Coverage per function — clean, finding, skipped-with-reason, or silent gap
Don't trust AI jargon.Trust a passing exploit.
Agents author exploit tests. They are never permitted to run them. A supervisor executes each attempt inside a disposable clone of your repository, records the receipt, and destroys the clone. No receipt, no claim.
Executed at your pinned commit. The exploit ships with the report, hash-verified.
Argued from source and gated. Marked on the finding, never counted as proven.
Whatever was not exhaustively checked is named, with the reason.
A report your team can act on, and your auditors can check.
A master report that is an index
Protocol, commit, audited scope, severity counts, and one row per finding with an explicit status — Reported, or Manual verification required. No padding.
One standalone file per finding
Submission-ready: summary, root cause, exact preconditions, attack path, impact, self-contained proof, source locations, recommended fix. No internal jargon.
A shipped PoC package
The exploit tests that actually ran, placed at the repository paths they were executed at and hash-verified against what the runner executed — so your engineers can run them.
Honesty artifacts
Defense-in-depth notes for fragilities that aren't findings yet, an incomplete-coverage appendix, and a retained internal set for your own review.
- Critical4
- High7
- Medium11
- Low9
- Native formats
- Immunefi
- Code4rena
- Cantina
- HackenProof
- Or your own template
Parsed and enforced section-by-section before writing.
Choose the engagement that fits your protocol.
Thorough
Launches, contests, protocol-wide coverageThe full engagement. Every discovery lane, convergence and depth iteration, machine-owned proof, one accountable verifier per finding, and the complete publication gate stack.
- 50–100 specialist agents across non-overlapping lanes
- Attack chains composed mechanically — matching preconditions, capabilities, capital and timing windows
- Per-phase checkpointing with full resume after any interruption
Frontier
Teams and researchers who need the unknownA differentiated hunt: structural risk graph, hotspot micro-context, competitive exploit-hypothesis sets, and extra depth lanes aimed at novel bug classes. Always ships submission-readiness QA.
- Extra depth lanes targeting bug classes with no existing detector
- Competitive exploit-hypothesis sets for contest and bounty work
- Scope auto-detection for contest formats; severity scoped to the bands your program rewards
Make security the first line,not the postmortem.
Engagement-based. Bring a repository, a commit hash, your program's rules and the severity bands you care about — we take scope up front and hand back attested proof.