Agent Skills: Building Secure Contracts Skill

Building Secure Contracts Skill

Overview

This skill applies systematic security analysis to smart contracts and secure API contracts. The core principle: every state mutation must be proven safe through invariant verification before an external call executes. It covers both EVM (Solidity) and Solana (Rust) ecosystems with platform-specific vulnerability patterns.

Vulnerability taxonomy: OpenSCV (94 classified security issues) Critical patterns: CEI, reentrancy guards, access modifiers, SafeMath equivalents Risk landscape: $1.8B+ in DeFi exploits Q3 2025 (access control: $953M, reentrancy: $420M)

When to Use

• Before deploying any smart contract to mainnet
• When auditing existing contracts for security vulnerabilities
• When reviewing API contracts for invariant violations
• When adding new entry points or external calls to existing contracts
• When upgrading proxy contracts (storage slot collision risk)
• When integrating oracles, flash loans, or third-party DeFi protocols

Iron Laws

1. NEVER make external calls before updating state — Checks-Effects-Interactions (CEI) is non-negotiable; any external call before state update is a reentrancy vector regardless of perceived safety.
2. NEVER assume access control is correct without reading every modifier — access control failures account for ~53% of 2024 DeFi losses; verify every onlyOwner, onlyRole, and custom guard.
3. NEVER trust integer arithmetic without explicit bounds checking — Solidity 0.8+ has native overflow protection but custom assembly, unchecked blocks, and Rust/Solana code require explicit verification.
4. ALWAYS enumerate all contract invariants before analysis — invariants are the ground truth for correctness; a violation is always a bug; document them in NatSpec before reviewing the implementation.
5. ALWAYS test reentrancy across full call chains, not just single functions — cross-function reentrancy (withdraw + transfer sharing state) is as dangerous as direct reentrancy.

Phase 1: Contract Reconnaissance

Goal: Map the attack surface before deep analysis.

Steps

1. Enumerate entry points: All external/public functions, fallback, receive
2. Identify state-mutating functions: Functions that modify storage
3. Map access control boundaries: Roles, modifiers, ownership checks
4. Catalog external calls: call(), transfer(), ERC20 hooks, interface calls
5. Identify trust boundaries: User input, oracle feeds, cross-contract calls

Output Format

## Contract Reconnaissance

### Entry Points

- [ ] `withdraw(uint256 amount)` — external, state-mutating, calls msg.sender
- [ ] `deposit()` — payable, updates balances mapping

### Access Control Map

- [ ] `onlyOwner`: [list of functions]
- [ ] `onlyRole(ADMIN_ROLE)`: [list of functions]
- [ ] No modifier (verify intent): [list of functions]

### External Calls

- [ ] `msg.sender.call{value: amount}("")` at withdraw():L45
- [ ] `token.transferFrom(...)` at deposit():L23

### Trust Boundaries

- [ ] User-supplied amount at withdraw():L40
- [ ] Oracle price feed at getPrice():L67 — manipulation risk

Phase 2: Reentrancy Analysis

Goal: Identify all reentrancy vectors (direct, cross-function, read-only).

Checks-Effects-Interactions Verification

For each function with external calls:

### Function: withdraw(uint256 amount)

#### CEI Order Analysis

- L40: CHECK — require(balances[msg.sender] >= amount) ✓
- L45: EXTERNAL CALL — msg.sender.call{value: amount}("") ← VIOLATION
- L48: EFFECT — balances[msg.sender] -= amount ← STATE AFTER CALL

**FINDING**: Classic reentrancy — balance updated after external call.
**Fix**: Move L48 before L45 (CEI pattern)
**Severity**: Critical

#### Fixed Pattern

```solidity
require(balances[msg.sender] >= amount);
balances[msg.sender] -= amount;  // Effect BEFORE external call
(bool success, ) = msg.sender.call{value: amount}("");
require(success);
```

Cross-Function Reentrancy Check

Identify shared state between functions that both make external calls:

### Shared State: balances mapping

- withdraw() reads + writes balances + makes external call
- emergencyWithdraw() reads + writes balances + makes external call
  **RISK**: Reentrancy from withdraw() into emergencyWithdraw() bypasses checks

Phase 3: Access Control Audit

Goal: Verify every state-mutating function has appropriate guards.

Access Control Checklist

For each function:

### Function Audit: updateTreasury(address newTreasury)

- [ ] Has access modifier? → NO ← FINDING: Missing onlyOwner
- [ ] Modifier verified in contract? → N/A (not present)
- [ ] Owner transferable safely? → N/A
- [ ] Time lock for critical changes? → NO

**Severity**: Critical — anyone can redirect protocol treasury
**Fix**: Add `onlyOwner` modifier and time-lock for parameter changes

Role Confusion Patterns

### Role Check: PAUSER_ROLE vs ADMIN_ROLE

- pause() requires: PAUSER_ROLE
- unpause() requires: PAUSER_ROLE (RISK: pauser can also unpause)
- grantRole() requires: ADMIN_ROLE

**Issue**: Pauser can unilaterally pause and unpause — should require separate roles
**Severity**: Medium

Phase 4: Integer Arithmetic Analysis

Goal: Identify overflow, underflow, precision loss, and rounding direction bugs.

Arithmetic Boundary Analysis

### Function: calculateReward(uint256 principal, uint256 rate)

- L88: `uint256 reward = principal * rate / 1e18`
  - Multiplication before division: OK (avoids precision loss)
  - Overflow check: principal \* rate could overflow if both > sqrt(uint256.max)
  - Rounding: truncates toward zero — check if favors protocol or user
  - `unchecked` block? → NO → Solidity 0.8+ protects this

### Unchecked Block Analysis

- L102-108: `unchecked { ... }`
  - Why unchecked? Check comment and verify mathematician's claim
  - Is the claimed impossibility of overflow actually proven?
  - [UNVERIFIED] claim: "amount < balance guarantees no underflow"

Phase 5: Invariant Verification

Goal: Identify and verify all contract-level invariants.

### Contract Invariants: LiquidityPool

1. **Solvency**: sum(balances) == address(this).balance — [VERIFIED L90]
2. **Total supply**: totalSupply == sum(all user shares) — [UNVERIFIED]
3. **Fee bound**: fee <= MAX_FEE (1000 bps) — [VERIFIED by require at L45]
4. **Non-zero denominator**: totalSupply > 0 before share calculation — [VIOLATED at L67, division-by-zero risk on first deposit]

### Invariant Violation Findings

**FINDING**: Invariant 4 violated — first depositor can cause division by zero

- Location: L67 `shares = amount * totalSupply / totalAssets`
- When: totalSupply == 0 on first deposit
- Impact: DoS attack on first deposit; protocol initialization blocked
- Fix: Handle zero totalSupply case separately with initial share ratio

Output: Security Report

# Security Report: [Contract Name]

## Summary

- Functions analyzed: N
- Findings: N (Critical: X, High: Y, Medium: Z, Low: W)
- Invariants verified: N of M
- CEI violations: N

## Critical Findings

### [F-01] Reentrancy in withdraw()

- Location: `src/Pool.sol:L45`
- Pattern: External call before state update (CEI violation)
- Impact: Complete fund drainage
- Fix: Apply CEI pattern — update state before external call
- 5 Whys: [root cause chain]

## Invariant Status

| Invariant                  | Status     | Evidence            |
| -------------------------- | ---------- | ------------------- |
| sum(balances) == balance   | VERIFIED   | L90 invariant check |
| totalSupply == sum(shares) | UNVERIFIED | No test coverage    |

## Recommendations

1. [Critical] Fix reentrancy in withdraw() before deployment
2. [High] Add reentrancy guard as defense-in-depth
3. [Medium] Add formal invariant tests via Foundry invariant suite

Integration with Agent-Studio

Recommended Workflow

1. Invoke audit-context-building for initial code reconnaissance
2. Invoke building-secure-contracts for contract-specific analysis
3. Feed findings into security-architect for threat modeling
4. Use static-analysis (Semgrep/CodeQL) for automated confirmation
5. Use medusa-security for fuzzing-based invariant testing

Complementary Skills

| Skill | Relationship | | ------------------------ | ---------------------------------------------------- | | audit-context-building | Builds initial mental model before contract analysis | | security-architect | Consumes findings for threat modeling and STRIDE | | static-analysis | Automated SAST confirmation of manual findings | | medusa-security | Fuzzing and property-based testing for invariants | | variant-analysis | Finds similar vulnerability patterns across codebase | | web3-expert | Solidity/Ethereum ecosystem expertise |

Anti-Patterns

| Anti-Pattern | Why It Fails | Correct Approach | | ----------------------------------------------- | -------------------------------------------------------------- | ----------------------------------------------------------- | | Auditing only the happy path | Reentrancy and access control bugs are invisible in happy path | Explicitly trace every error path and external call | | Trusting function name for access control | onlyAdmin() might not check the actual admin role | Read the modifier implementation, not just its name | | Assuming Solidity 0.8 prevents all integer bugs | unchecked blocks, assembly, and casting bypass protection | Audit all unchecked blocks and type casts explicitly | | Skipping cross-function reentrancy | Cross-function reentrancy bypasses single-function guards | Map shared state across ALL functions making external calls | | Leaving invariants implicit | Unwritten invariants are unverified risks | Document every invariant in NatSpec before analysis |

Memory Protocol

Before starting: Check .claude/context/memory/learnings.md for prior contract audits of the same protocol or token standard.

During analysis: Write incremental findings to context report as discovered. Do not wait until the end.

After completion: Record key findings and patterns to .claude/context/memory/learnings.md. Record architecture decisions (CEI enforcement patterns, invariant frameworks) to decisions.md.