web3-audit – Security skill for AI agents

WEB3 SMART CONTRACT AUDIT

10 bug classes. Pre-dive kill signals. Foundry PoC template. Real paid examples.

---

PRE-DIVE KILL SIGNALS (check BEFORE any code review)

> ZKsync lesson: $322M TVL + OZ audit + 750K LOC + 5 sessions = 0 findings. Large well-audited bridges are extremely hard.

1. TVL < $500K → max payout capped too low for effort

2. 2+ top-tier audits (Halborn, ToB, Cyfrin, OpenZeppelin) on simple protocol → bugs already found

3. Protocol < 500 lines, single A→B→C flow → minimal attack surface

4. Formula: max_realistic_payout = min(10% × TVL, program_cap) — if < $10K, skip

Soft kill: OZ/ToB/Cyfrin audit on current version + codebase > 500K LOC → expect 40+ hours for maybe 1 finding. Only proceed if bounty floor > $50K AND you have protocol-specific expertise.

Target scoring (go if >= 6/10):

• TVL > $10M: +2
• Immunefi program with Critical >= $50K: +2
• No top-tier audit on current version: +2
• < 30 days since deploy: +1
• Protocol you've hunted before: +1
• Source code + natspec comments: +1
• Upgradeable proxies: +1

---

THE ONE RULE

> "Read ALL sibling functions. If vote() has a modifier, check poke(), reset(), harvest(). The missing modifier on the sibling IS the bug."

This single rule explains 19% of all Critical findings.

---

1. ACCOUNTING STATE DESYNCHRONIZATION

> #1 Critical bug class — 28% of all Criticals on Immunefi.

What It Is

Two state variables supposed to stay in sync. One code path updates A but forgets B. Later code reads both and makes decisions based on stale B.

Real Value = A - B
If A updated but B isn't → Real Value appears larger → phantom value

Root Cause Patterns

Variant 1: Phantom Yield (Yeet protocol — 35 duplicate reports)

function startUnstake(uint256 amount) external {
    totalSupply -= amount;  // decremented BEFORE transfer
    // aToken.balanceOf(this) still reflects old value
    // yieldAmount = aToken.balanceOf - totalSupply = phantom yield
}

Variant 2: Fast Path Skips State Update (Alchemix V3)

function claimRedemption(uint256 tokenId) external {
    if (transmuter.balance >= amount) {
        transmuter.transfer(user, amount);
        _burn(tokenId);
        return;  // EARLY RETURN — cumulativeEarmarked, _redemptionWeight, totalDebt never updated
    }
    // Slow path: updates all state vars correctly
    alchemist.redeem(...);
}

Variant 3: Update Happens in Wrong Order (Alchemix)

function deposit(uint256 amount) external {
    _shares = (amount * totalShares) / totalAssets;  // calculated BEFORE deposit
    totalAssets += amount;   // assets added AFTER shares calculated → wrong rate
}

Grep Patterns

# Find all accounting variables
grep -rn "totalSupply\|totalShares\|totalAssets\|totalDebt\|cumulativeReward\|rewardPerShare" contracts/

# Find all early returns in claim/redeem functions
grep -rn "\breturn\b" contracts/ -B3 | grep -B3 "if\b"

# For each early return: which state updates in normal path are skipped?

---

2. ACCESS CONTROL

> #2 Critical — 19% of Criticals. $953M lost in 2024 alone.

Variant 1: Missing Modifier on Sibling Function

function vote(uint256 tokenId) external onlyNewEpoch(tokenId) {  // guarded
function reset(uint256 tokenId) external onlyNewEpoch(tokenId) { // guarded
function poke(uint256 tokenId) external {                         // NO GUARD → infinite FLUX inflation
}

Variant 2: Wrong Check (Existence vs Ownership)

function split(uint256 tokenId, uint256 amount) external {
    _requireOwned(tokenId);  // checks if token EXISTS, not if caller OWNS it
    _burn(tokenId);
    _mint(msg.sender, amount);  // attacker steals tokens they don't own
}

Variant 3: Silent Modifier (if vs require)

// VULNERABLE — non-admin silently gets through:
modifier onlyAdmin() {
    if (msg.sender == admin) {
        _;  // body only executes for admin, but non-admin doesn't revert
    }
}
// CORRECT: require(msg.sender == admin, "Not admin"); _;

Variant 4: Uninitialized Proxy

function initialize(address _owner) public {  // MISSING: initializer modifier
    owner = _owner;  // anyone can call → become owner
}
// Fix: constructor() { _disableInitializers(); }

Grep Patterns

# Find sibling function families — do ALL have the same modifier set?
grep -rn "function vote\|function poke\|function reset\|function update\|function claim\|function harvest" contracts/ -A2

# Ownership check: existence vs ownership?
grep -rn "_requireOwned\|ownerOf\|_isApprovedOrOwner\|_checkAuthorized" contracts/ -B5

# Silent modifiers
grep -rn "modifier\b" contracts/ -A8 | grep -B3 "if (" | grep -v "require\|revert"

# Uninitialized initializer
grep -rn "function initialize\b" contracts/ -A3
grep -rn "_disableInitializers()" contracts/

Real Paid Examples

| Protocol | Payout | Bug |

|---|---|---|

| Wormhole | $10M | Uninitialized UUPS proxy → anyone calls initialize() |

| ZeroLend | n/a | split() uses existence check, not ownership check |

| Alchemix | n/a | poke() missing onlyNewEpoch → infinite FLUX inflation |

| Parity | $150M frozen | No access control on initWallet() in library |

---

3. INCOMPLETE CODE PATH

> #3 Critical — 17% of Criticals.

The Function Family Comparison Test

1. List all state changes in function A (deposit/place/create)
2. List all state changes in function B (withdraw/update/cancel)
3. For each state change in A: does B have the corresponding reverse?
4. For each token transfer in A: does B have the corresponding refund?
If A does X but B doesn't do the reverse of X → BUG.

Variant 1: Update Function Missing Refund (ThunderNFT)

function place_order(OrderInput calldata order) external {
    token.safeTransferFrom(msg.sender, address(this), order.price);  // takes tokens
    orders[orderId] = order;
}
function update_order(OrderInput calldata updatedOrder) external {
    // BUG: NO REFUND for sell orders when price decreases → tokens permanently stuck
    orders[orderId] = updatedOrder;
}

Variant 2: Partial Fill Token Stuck (Plume)

function swapForETH(uint256 amountIn) external {
    token.safeTransferFrom(msg.sender, address(this), amountIn);
    uint256 filled = dex.swap(amountIn);  // partial fill possible
    _refundExcessEth(amountIn - filled);  // BUG: refunds ETH only, not ERC20
}

Variant 3: mint() Bypasses Check That deposit() Has (MetaPool)

function deposit(uint256 assets, address receiver) public override {
    shares = _deposit(assets, receiver);  // includes receipt validation
}
function mint(uint256 shares, address receiver) public override {
    assets = convertToAssets(shares);
    _mint(receiver, shares);  // MISSING: _deposit() validation → mints without receiving assets
}

Grep Patterns

grep -rn "function place_\|function create_\|function add_\|function open_" contracts/ -A5
grep -rn "function update_\|function modify_\|function cancel_" contracts/ -A5
grep -rn "safeApprove\b" contracts/    # safeApprove without zero-reset before
grep -rn "delete\b" contracts/ -B5 -A5  # delete before operation completes
grep -rn "function deposit\|function mint\|function withdraw\|function redeem" contracts/ -A10

---

4. OFF-BY-ONE & BOUNDARY CONDITIONS

> #4 High — 22% of Highs. Single character change. Massive impact.

Root Cause

// VeChain Stargate — post-exit reward drain:
function _claimableDelegationPeriods(address delegator) internal view returns (uint256) {
    if (endPeriod > nextClaimablePeriod) {  // BUG: should be >=
        return 0;  // exited users get nothing
    }
    return nextClaimablePeriod - lastClaimedPeriod;  // rewards for period AFTER exit
}

Mental Test for Every Comparison

> For every if (A > B): "What happens when A == B?" Is that correct?

6 Boundary Locations to Check

1. Period/Epoch boundaries: > vs >= at period end

2. Time-based locks: does block.timestamp == deadline lock or unlock?

3. Loop break conditions: break with > vs >=

4. Array index boundaries: i <= array.length (should be i < array.length)

5. Amount/balance boundaries: >= amount allows exact full withdrawal?

6. Rounding/precision: can any input produce 0 output that should be non-zero?

Grep Patterns

# Boundaries in comparisons
grep -rn "Period\|Epoch\|Round\|Deadline\|period\|epoch\|deadline" contracts/ -A3 | grep "[<>][^=]"

# Loop breaks
grep -rn "\bbreak\b" contracts/ -B10

# Off-by-one in array access
grep -rn "\.length\s*-\s*1\|i\s*<=\s*.*\.length\b" contracts/

---

5. ORACLE / PRICE MANIPULATION

> 12% of all reports. Largest individual payouts. $117M Mango, $70M Curve.

Bug A: Missing Staleness Check (most common)

// VULNERABLE:
(, int256 price,,,) = priceFeed.latestRoundData();
return uint256(price);  // If Chainlink node goes down, stal

… (full content is in the download package)