kraitAudit your contracts with
your IDE’s AI.
Krait runs 845 security checks against your contracts using Claude Code, Cursor, Windsurf, or Codex. Each check generates a prompt grounded in real Solodit exploits — Krait parses the verdict for you. Free.
From 12% to 100% — eight versions, every miss tracked.
We ran Krait blind against 40 past Code4rena, Sherlock, and CodeHawks contests — real protocols with known vulnerabilities.
Methodology Evolution
Our #1 goal: eliminate false positives
Every AI audit tool can find bugs. The hard part is not crying wolf. A report full of false positives wastes your team's time, erodes trust, and buries the real vulnerabilities that actually matter. That's why we built the entire pipeline around FP elimination — from the Critic agent's 8 Kill Gates to post-mortem analysis of every false positive across 40 shadow audits.
GoodEntry
5/14 findings on UniV3 derivatives — slot0 manipulation, flash loan initiator, V3Proxy routing
Munchables
Found both HIGH findings, 0 false positives
Arcade
Caught voting power desync on multiplier change — exact Pashov vector
A report your auditor will read.
Whether you use the web assessment or the CLI agent, you get a branded, professional report showing exactly where you stand.
kraitWhat's in your report
Overall security posture score with pass/fail/skip breakdown
Unresolved critical & high severity findings requiring immediate attention
High-level patterns we observe in your security architecture — like a senior auditor's executive summary
What you're doing right — documented strengths that demonstrate security maturity to investors and partners
Every check organized by section with pass rates and expandable details
Tool outputs, code references, and manual review notes for every assessed check
How it works — and why it takes 10 minutes, not 10 hours.
Every check generates a tailored AI prompt. Run it in your IDE. Paste the response back. Krait auto-parses the verdict. No setup, no install, no API key.
Tell Krait what you're building
Pick from 39 protocol types. Configure chain, admin model, oracle usage. Krait loads the right domain primer with protocol-specific attack patterns and relevant Solodit findings.
Every check generates a tailored AI prompt
Each check has a "Verify with AI" button that generates a prompt with real Solodit exploits, code patterns to search for, and what secure code looks like. Copy it into Claude Code, Cursor, Windsurf, or Codex. Paste the response back — Krait auto-detects the verdict, extracts file:line references, and sets the status.
You are a security auditor specializing in Lending protocols. Analyze for: oracle price freshness... ### Real exploits from production audits - [HIGH] Oracle staleness in Aave v3...
Professional report you can share with anyone
Get a readiness score, ranked risks by severity, strengths identified, and every finding linked to its source audit. Export as Markdown, share a URL, or bring it to an auditor. Know exactly what to fix before your audit.
Reentrancy on Vault.withdraw() via fee callback
External call to feeRecipient precedes shares[msg.sender] -= amount. Same pattern as the SushiMiso incident on Solodit.
A 4-stage adversarial pipeline
that disproves its own findings.
Most "AI audit tools" run a single LLM pass and dump every possible issue — burying real bugs under mountains of false positives. Krait's 4-stage adversarial pipeline is purpose-built to eliminate noise: agents detect, reason, disprove, and rank — achieving 95%+ fewer false positives than single-pass approaches.
Detector
Casts a wide net. 5 parallel workers scan every file using 101 detection heuristics, Feynman interrogation (9 question categories), and 140 vulnerability patterns.
Three-pass analysis · Per-file + cross-contract
Reasoner
Builds concrete exploitation proofs. Must construct a 3-step attack: attacker does X → causes Y → results in Z loss. No hand-waving allowed.
Only provably exploitable findings survive
Critic
Devil's advocate with 8 Kill Gates. Actively tries to DISPROVE each finding. Checks for existing mitigations, safe patterns, and 8 systematic false-positive rules.
Catches 95%+ of false positives
Ranker
Composite scoring: detector (20%) + reasoner (30%) + critic (50%). Threshold filtering removes remaining noise. Only validated findings survive.
Score ≥ 40/100 to pass · Severity-ranked output
Six sources behind every check.
Krait's knowledge comes from 6 distinct data sources, each curated, validated, and continuously refined through a 40-contest shadow audit feedback loop.
140 Vulnerability Patterns
Hand-crafted YAML
Each pattern contains detection strategy, real exploit examples with vulnerable + fixed code, false-positive notes, and confidence levels. Covers Solidity, Rust/Solana, AI/Red-Team, and Web2.
49K+ Solodit Findings
World's largest audit database
Real vulnerabilities from Aave, Uniswap, Compound, MakerDAO, and hundreds more. Fetched via API, categorized by protocol type, mapped to attack vectors. Rate-limited with intelligent caching.
40 Shadow Audits
C4 / Sherlock / CodeHawks
Every new methodology version is blind-tested against real contest results. Post-mortem analysis of every miss. Patterns and heuristics updated. 8 methodology versions from 12% to 100% precision.
333+ Protocol Primers
7 domain-specific guides
DEX/AMM (20 attack patterns), Lending (15), Staking (15), NFT/GameFi (17), Bridge (12), Proxy/Upgrades (15), Wallet/AA (12). Each distilled from hundreds of verified audit findings.
101 Detection Heuristics
Feynman Interrogation System
9 question categories (Purpose, Ordering, Consistency, Assumptions, Boundaries, Returns, External Calls, Protocol Integration, Overrides) with 40+ sub-questions for systematic code interrogation.
8 Kill Gates
Systematic FP elimination
First-depositor, fee-on-transfer, fork-inherited, admin-gated, best-practice, known issues, privilege escalation, acknowledged issues. Each gate is a learned filter from shadow audit false positives.
Continuous Learning Loop
Blind audit a real contest
Compare against actual results
Analyze every miss & false positive
Extract new patterns & heuristics
Update pipeline, test on next contest
This loop has run 40 times across 8 methodology versions. v1 started at 12% precision. v8 reached 100% precision with zero false positives across 5 contests.
“We built Krait because half our audit hours go to bugs the team could have found in a checklist. Now they can.”