Static engine on the corpus
of 143 contracts101 of 143 labelled targets matched (70.6%)
12 bug classes. An alarm outside the labelled class counts against precision, which is why this is recall and not accuracy.
Open the measurementQuick scan
One paste. Auto-detect route. Review the report.
VegaSec runs read-only checks on public sources and returns one report: what we found, how strong the evidence is, and what we could not check. No wallet, no login, no promise that a target is safe.
Free, no wallet, read-only, shareable report URL.
Buyer-ready sample report
Pre-Launch Token Watchlist
This is an evidence level, not a safety verdict.
Could not check
Repository Audit
8 of 9
The 0 to 100 score lives inside the report, after the evidence and the unknown items, with the formula published. It is derived, not the answer, and it does not include what the scan could not check.
Example report, generated from a public sample target.
Measured, not claimed
Four figures this product can prove today. Each one links to the page or the stored artifact it was read from, and each one names the sample it was measured on.
One artifact per figure
Static engine on the corpus
of 143 contracts101 of 143 labelled targets matched (70.6%)
12 bug classes. An alarm outside the labelled class counts against precision, which is why this is recall and not accuracy.
Open the measurementReviewer roles on the same contracts
of 143 contracts118 of 137 role targets found (86.1%)
9 roles, each measured on its own class. A role that found nothing of its class is not counted as zero.
Open the measurementContracts run in the public lane
in the lane10 contracts with a stored signed report
Bug class from the documented corpus, counts from the stored report. Static analysis: the class is matched, not exploited.
Open a stored reportArtifact anyone can verify
key vsig-2026-0910 of 10 batch reports signed
The certificate prints only the fields the signature covered: target, engine versions, severity counts, date and limitations.
Open a certificateEngine and reviewer figures come from one public corpus (SmartBugs-Curated, MIT, commit 230e649), measured on the same 143 contracts and reconciled by test against the raw artifacts. The batch figure comes from the contracts run in the public lane, each with a stored report; those reports carry static findings only, with no operator judgment attached. Static analysis matches a documented class: it does not reproduce an exploit.
Our measurement
Two numbers describe what VegaSec can and cannot see, both measured on the same public corpus, plus the live index of the contracts we have already scanned. Each claim carries its evidence level, its sample and what it does not say.
One artifact per claim
Deterministic engine on the corpus
EVIDENCE: STRONG101 of 143 labelled expectations matched (70.6%)
Sample: 143 contracts, 12 classes, SmartBugs-Curated (MIT, commit 230e649123).
Outside scope: 42 expectations were not matched and 200 flags fell outside the labelled class, which counts as an alarm even when the finding is real (precision 33.6%). The classes centralization, logic, uninitialized have no label in the corpus, so their recall is not measured.
143 of 143 in scope, 0 outside
Hunter roles on the same corpus
EVIDENCE: PARTIAL118 of 137 role targets found with the quote inside the cited lines (86.1%)
9 hunter roles, 1,277 of 1,287 calls answered, the same 143 contracts.
Outside scope: 9 fixtures sit in classes with no role (front-running, timestamp-dependence), 4 targets went unanswered, and 2 roles found nothing of their own class (hunt_oracle_price_manipulation, hunt_signature_replay). Fixture precision is 56.1%.
134 of 143 in scope, 13 outside
Our findings judged by a second reader
EVIDENCE: PARTIAL34 of 37 findings confirmed by a second reader (3 partial, 0 wrong)
Adversarial verifier checked 37 items: it caught 3/3 deviations and wrongly flagged 14/34 correct findings.
Outside scope: a partial verdict is not a confirmation, and the verifier checks the reader instead of labelling the finding, so confirmation is a floor: 14/34 correct findings were flagged by it.
37 of 37 in scope, 3 outside
Contracts we have already scanned
EVIDENCE: PARTIAL50 contracts in the public index
Live read of the same public index that /contracts renders, newest first, capped at 50.
Outside scope: contracts we never scanned and any protocol whose scan is not listed. The count describes the index, not the market.
50 of 50 in scope, 2 outside
Sources: docs/benchmarks/2026-09-26-smartbugs-143.json and docs/benchmarks/2026-09-26-role-scorecard-corpus143-r4.rescored.json, reconciled by test against the raw artifacts.
Validated cases
Every card below is a public contract this engine scanned. The bug class comes from the documented corpus, the counts and the findings come from the stored artifact, and the link opens the stored report. Static analysis only: it matches a documented class, it does not reproduce an exploit.
12 contracts with a stored report
Pick a contract to open its stored audit.
12 cases shown
Documented corpus
reentrancy-simple-dao
What the scan marked
Severity split
Audit pipeline
Replay of the audit pipeline the engine runs: intake, source acquisition, toolchain validation, static analysis, finding normalization, exploitability review, remediation draft, operator review and finalization. The finding counts come from the stored report. The per-stage budget is the replay pacing used on this page, not live worker timing.
Ready to replay
Replay of stored data. No live scan runs on this page.
These stored reports carry no operator review artifact, so the pipeline stops at the automated report. A human review is a separate engagement, quoted per target.
Free, read-only, no login and no wallet. A human review is quoted per target.
The problem, with the source named
Two public series, each with the report it comes from and the link to read it. No fear number of our own.
TRM Labs, 2026 Crypto Crime Report
$2.87B stolen across about 150 hacks and exploits in 2025
70% of that value sits in five events, and the Bybit breach alone accounts for $1.46B (51%). Source: TRM Labs, 2026 Crypto Crime Report.
Read the reportImmunefi, Ecosystem Vulnerability Scoreboard
DeFi protocol losses fell from $2.62B in 2022 to $534M in 2024, then returned to $680M in 2025
The median loss per incident fell 75%, from $6M to $1.5M, while the number of incidents rose. Source: Immunefi, Ecosystem Vulnerability Scoreboard.
Read the reportThe path, step by step
Four steps and the door each one opens. The scan is free and public. The rest is scoped and quoted in writing before anything runs.
Pick a step
01 · Free scan
No login, no walletPaste a contract, a token, a repository or an explorer URL. The engine reads public sources, writes the report and leaves it at a shareable URL. A missing source fails closed to UNKNOWN instead of being filled in.
Read-only and public. This step is free and carries no operator approval.
The scan stays free. The paid steps quote before they run, and the certificate is an artifact served today.
A clean left column never cancels the right one.
Methodology
VegaSec keeps the report model consistent: public Scan creates the baseline, Private Review adds operator approval, and Continuous Watch checks whether code or contract state changed after the baseline. Score, confidence, evidence, limitations and next action stay visible across all three.
Public sources VegaSec reads
Every signal comes from a public provider we name. No paid data resold, no opaque "intelligence feed", no scraping behind login walls.
All keys are server-side only
DexScreener
DEXToken + Pair market data
Public, key-less pair lookup. Liquidity / holder posture / trading signal.
GitHub public API
GITRepository metadata + commits
Default branch, latest commit SHA, license, README, CI signal, security policy.
Etherscan v2 · Basescan
SRCVerified Solidity source
Acquisition path for slither + semgrep static analysis. Key never exposed to the browser.
Public chain RPC
RPCNFT interface probes
ERC-721 / ERC-1155 detection, tokenURI inspection, owner / proxy state. Anonymous read.
VegaSec does not state that a target is safe and does not give investment advice. Every result is a snapshot of public data at scan time. A missing source fails closed to UNKNOWN.