What an AI Auditor Can and Can't Find

An AI audit can produce useful findings and still miss a serious bug. Its limits depend on the model, the code and dependencies available to it, the review scope, and how its hypotheses are tested. Understanding those limits helps you decide what evidence to request and where expert review is still needed.

What AI Auditing Is Good At

AI review can help investigate vulnerabilities with well-understood mechanics, especially when the challenge is tracing how they apply to a particular codebase:

• Known vulnerability classes with many variations. Reentrancy (including read-only reentrancy), oracle manipulation, access-control gaps, precision loss, and unsafe token integrations give a reasoning engine concrete mechanisms to investigate. No listed class is guaranteed to be detected.
• Repeatable checks. Automation can repeat a configured review, but equal scrutiny of every function is not guaranteed. Inspect the scope, completed stages, and skipped files or tools.
• Repeatable review as code changes. Timing depends on repository size, scope, compilation, and the hypotheses investigated. Firepan's free surface scan is designed for quick initial triage; a HOUND deep audit is a separate, asynchronous workflow with variable duration.
• Surfacing suspicious paths with evidence. An unconfirmed exploit hypothesis can still identify a path worth investigating, provided the report distinguishes that hypothesis from a validated finding.

These checks can support initial triage and repeated review. They do not clear the codebase of known bug classes, and an empty report is not evidence of safety.

What AI Auditing Struggles With

The weaknesses are real and worth naming plainly:

• Novel, protocol-specific business logic. A flaw may depend on a protocol's unique economic design. If intended behavior is missing from the review inputs, a code-only reviewer can lack the context needed to recognize it. Include specifications and invariants when investigating logic errors; a model may still misunderstand them.
• Economic and game-theoretic design flaws. Whether an incentive can be gamed, or a liquidation mechanism is stable under stress, is a modeling question that often requires reasoning about markets and adversaries beyond the code itself.
• Cross-protocol and composability risk. A contract can be correct in isolation and unsafe once composed with a specific external protocol whose behavior isn't in the reviewed repository.
• Intent and trust decisions. AI can flag that an owner can mint or upgrade — that's centralization risk — but whether that authority is acceptable is a judgment call about the project's trust model, not a code fact.
• Confident wrong answers. Like any AI system, an auditor can produce plausible but incorrect reasoning. This is why serious AI auditing outputs scoped evidence a human can verify, rather than unexplained verdicts — and why "automatically catches everything" is a claim to distrust from any vendor.

Why Human Review Still Matters

AI and human reviewers can miss different issues, and sometimes the same ones. Expert reviewers add context about intent, economics, and design while checking machine-generated claims. Techniques like formal verification can prove specified properties within a model and its assumptions; they do not prove that an entire deployed system is safe.

Combine repeatable automated checks with tests and focused expert review as code changes. That's the role of Firepan's AI-powered smart contract audit: to add evidence and repeat checks, with coverage bounded by the repository and completed analysis.

Frequently Asked Questions

Q: Can AI audit smart contracts reliably?

A: It can help find vulnerabilities, but reliability must be assessed for the actual tool, code, scope, and validation method. It can miss known bug classes as well as novel business-logic flaws. Request reproducible evidence and use it alongside testing and expert review.

---

Q: What can't AI auditing find?

A: It struggles with protocol-specific business-logic bugs where the intended behavior isn't expressed in code, economic and game-theoretic design flaws, composability risk involving external protocols, and trust-model judgments. It can also produce confident but incorrect reasoning, so findings should be verifiable evidence, not unexplained verdicts.

---

Q: Is AI auditing better than a human auditor?

A: There is no universal ranking. Compare results for the relevant codebase and review scope, including missed issues and false positives. Combining automated checks with expert review can add evidence, but it does not guarantee complete coverage.

---

Q: Should I skip a manual audit if I use AI auditing?

A: Do not use an AI report as a reason to skip expert review of a high-value protocol. Use repeatable automated checks alongside manual review and testing; none of these methods guarantees that the code is vulnerability-free.

---

Q: How do I know if an AI auditing claim is trustworthy?

A: Look for scoped evidence — the file, path, and conditions behind each finding — rather than an unexplained pass/fail, and be skeptical of any tool that claims to automatically catch every vulnerability. Honest coverage varies by language, code, evidence, and scan type.

Where to Go Next

See the vulnerability database, or start a free public-repository surface scan for initial triage. A surface scan uses deterministic detectors with optional LLM verification; it is not a HOUND deep audit.

Sources

• Solidity security considerations: security pitfalls and review recommendations.
• Firepan audit scope and workflow: the distinction between surface-scan timing and asynchronous deep audits.