Depth Is the Product.
Most scanners hand you a list of findings, sorted by severity. SaferICO composes them into the chain that uses them — and ships the evidence for every link it claims.
A finding is a line. An attack is a combination of them — and that is what we look for.
Each layer sees what the one above it cannot.
The order is not presentational. Every layer consumes what the previous one produced, and each is defined as much by its blind spot as by its catch. Those blind spots are printed here for the same reason they are printed in the report: a tool that hides its reach is a tool you cannot calibrate.
Detect
201 detectors over Solidity source: unchecked returns, reentrancy shapes, mint and sweep paths, owner powers, proxy and upgrade paths, and a simulated buy and sell that catches a honeypot before you are in one. The engine runs on your machine — an unreleased contract is never uploaded to read it.
Corroborate
Source is a claim; deployed bytecode is the fact. This layer reads the dispatcher for the selectors a contract really exposes, names raw calldata against a signature corpus, subtracts library code so only the hand-written lines remain — the lines an auditor should read first — and reports where an address's funds came from, each answer carrying the provenance of the list that produced it.
Compose
Each chain carries a question — “can I make one share so expensive that a real deposit buys zero of them?” — a list of preconditions, and the weight each one carries. A chain is emitted only when the scan actually proved its required preconditions, and it ships the evidence for every one. If the code does not support the chain, no chain appears.
Explain
The model's work begins only after layer 03 has established a chain: explain that chain, in the reader's own language, at the level a defender needs to close it. It is never asked for scenarios and it cannot add one — because a confident, unfalsifiable list of invented attacks buries the single real finding underneath it.
Record
Every finding with its severity, its evidence and its fix. A public archive of the reports already delivered. A badge that keeps what was read live on-chain apart from what was merely reported, and dates the second. And monitoring that keeps watching the address after the report is signed, because a contract can change owner the week after it passes.
Scale, depth and restraint are three different claims.
Each number below is countable, and each answers a different question. None of them is a rounded marketing figure, and the third one is a limit we chose.
A public record across BNB Chain, Ethereum, Base, Polygon, Arbitrum, Optimism and Avalanche, 2021–2025. Every entry links the auditor’s own published PDF — not a copy hosted here, and not a summary of one.
Count them yourself →The detectors are the vocabulary; the chains are the sentences they can form. Each names its preconditions and the weight each carries. Twelve cite a documented incident that followed the same shape — the other nine state that no comparable one exists, rather than reaching for a plausible name.
Read all 21 →Every chain is composed in code from detector output and ships the evidence for each precondition it claims. The AI explains a chain that already exists. It has never been able to add one.
Two accepted weaknesses. One empty pool.
This is what layer 03 does with layer 01's output. Neither half below is critical on its own — which is exactly why a findings list, sorted by severity, hides it. The pattern is not hypothetical: Euler Finance lost $197M to accounting that could be moved the same way.
The composition High Precedent: Euler
What each layer contributed
01 Detect — the evidence. A share-manipulation detector and a rounding detector fired, and the contract can receive value directly. Three findings, none of them critical, sitting in different parts of the list.
02 Corroborate — the shape. Nothing named MIN_,
deadShares or initialMint appears anywhere in the function
table. Reading the deployed selectors is what separates absent from
unverified — the dead-share defence is missing, not merely unread.
03 Compose — the chain. Required preconditions satisfied; the chain is emitted with the evidence for each attached, and its confidence states exactly which share of them the scan could prove.
04 Explain — the walkthrough. The four steps above, in the reader's own language, at the level needed to close the path.
Five layers, three tools and a record.
The layers above are not a diagram of something abstract. Each one is a tool you can run today, and each answers a question the others cannot. Start anywhere — the output of one is the input of the next.
25 EVM mainnets, and the reason each one is on the list.
A network is added only after two checks, both measured on the day rather than read off a chain list: two or more public RPC endpoints answered with the right chain id and a block seconds old, and the chain is served by the explorer API key we already hold, so verified source can genuinely be fetched there. A chain with a live RPC and no explorer API gives half a scan — and half a scan reported as a scan is worse than none.
Reading a chain and building on one are different capabilities, and the list above says which is which rather than averaging them into one number. The detector set is identical everywhere — there are no chain-specific detectors, and a page implying otherwise would be a claim we cannot support. What changes by chain is the explorer, the endpoints, the router the buy/sell simulation uses, and what has already gone wrong there.
The Suite is testing. The audit is judgement.
They are sold separately because they answer different questions. One runs on every commit; the other is a named human on an engagement.
SaferICO Security Suite
Five layers, on every push. The detector engine runs inside your own CI runner, so your repository never leaves it — only counts and rule ids are posted back.
- Runs continuously, not once
- Every chain ships the evidence for each precondition
- Scales with your commits, not with our headcount
Manual Audit
A named auditor on economic design, governance and intent — the questions no pipeline can answer, because they are not questions about code.
- Separate from the Suite and from any subscription
- Starts from the Suite's output, not from zero
- Priced per engagement
What this cannot tell you.
Nothing here is executed
No transaction is sent, and no working exploit ships with this product. Layer 03 establishes that the preconditions for an attack are present in the code — it does not establish that the attack would succeed against live balances, liquidity and key holders. That distinction is carried in every result rather than left to be discovered.
A clean run is not a proof
It means no finding surfaced in what was tested, and no chain's preconditions were met. The honest claim is coverage, never safety.
Code is not the whole system
The Suite reads what is deployed. It cannot tell you who holds the keys, how they are stored, or what a privileged operator decides on a bad day.
Economics sit outside
A contract can be flawless and the token design still fail. Incentives, emissions and governance are the audit's work, not the pipeline's.
The seven questions this page should have answered.
Does anything get executed against my contract?
No. No transaction is sent and no working exploit ships with this product. Layer 03 establishes that the preconditions for an attack are present in the code; it does not establish that the attack would succeed against live balances, liquidity and key holders. That distinction is printed in every result rather than left to be discovered.
Does my source code leave my machine?
The 201-detector engine runs in your own browser, and in CI it runs inside your own runner. An unreleased contract is never uploaded to read it. What the browser fetches is public: verified source from the block explorer, and public RPC reads.
From CI, what is posted back is the repository name, ref, commit and actor, a file count, the severity counts and the score, and each finding's rule id with its file and line. Never the source, and never the evidence strings — those quote your code, so they stay on your runner.
What is the difference between the Suite and a manual audit?
They answer different questions and are sold separately. The Suite is testing: five layers, on every push, priced as a subscription. A manual audit is judgement — a named auditor on economic design, governance and intent, questions no pipeline can answer because they are not questions about code. An audit starts from the Suite's output rather than from zero.
Is the AI writing the attack chains?
It has never been able to. Every chain is composed in code from detector evidence and ships the proof for each precondition it claims. The model's work begins only after a chain already exists: explain that chain, in the reader's language, at the level a defender needs to close it. It is never asked for scenarios, because a confident, unfalsifiable list of invented attacks buries the one real finding underneath it.
What does a clean result actually mean?
That no finding surfaced in what was tested, and no chain's preconditions were met. It is coverage, not safety. A report that promises safety is a report to distrust — so no result here uses the word.
Which networks does it run on?
25 EVM mainnets for scanning, monitoring and the badge. Every one of them cleared two measured checks: two or more public RPC endpoints answering with the right chain id and a block seconds old, and an explorer that serves verified source on the key we already hold. Building and deploying is priced on 7 of them. Manual audits are not limited to the list at all — a person reading Solidity needs no integration.
Can I run it on every commit?
Yes — that is what the CI action is for. It runs the same detector set inside your runner and fails the build on a threshold you choose. Watchtower is the other half: it keeps reading the deployed address after the report is signed, because a contract can change owner the week after it passes.
Run the Suite on your contract.
Five layers, one report, every chain with its evidence attached. If nothing composes, the report says exactly that — and shows you what it tested to find out.