The Next Frontier: Using AI Agents to Build Custom Slither Detectors

Description and use

Исследователь Web3 безопасности разрабатывает пользовательские детриторы Slither для контрачей CEX интерфейсов. Использование LLM для генерации scaffold-

Problem

Manual development of production-ready Slither detectors takes 2–3 days per new threat vector, and the threat landscape evolves faster than any single researcher can keep up with, creating a bottleneck in Web3 security research.

Outcome

AI-assisted detector generation reduced development time from 2–3 days to 4–6 hours per detector. LLM-based false positive filtering reduced manual review time by 60–70%. The researcher emphasizes that AI is an accelerator, not a replacement, and that every line of AI-generated code still requires human review.

Workflow

• Run Slither on a contract and export results to JSON
• Feed the JSON output to an LLM with a request to analyze findings and identify likely false positives based on context
• Use an LLM to generate a working detector scaffold starting from a blank file
• Refine the generated code, addressing false positive logic, context awareness, and edge cases
• Test the AI-generated detector on known vulnerable contracts before production use
• Log confirmed or rejected AI suggestions to build a dataset for future fine-tuning

Evidence

Writing one production-ready detector takes 2–3 days of focused effort: studying the vulnerability pattern, implementing the logic, testing against known contracts, eliminating false positives, and wiring it into CI/CD.

Source

Instead of starting from a blank file, I now use LLMs to generate a working scaffold. The result is typically 70% of a working detector.

Source

In my experiments, this approach has reduced manual review time by 60–70%.

Source

I once asked an LLM to write a reentrancy detector. It produced code that looked correct at first glance — but it was checking for an old syntax that is no longer used in modern Solidity. The detector would have missed every real vulnerability in modern codebases.

Source

Each file format has its own attack surface. Input validation at the plugin boundary will be crucial as Hermes handles more untrusted outputs.

Source

A detector can flag 'admin function without timelock' as a vulnerability. But what if the protocol intentionally uses a fast admin path for emergency pauses? What if the 'missing' access control is actually enforced off-chain through a multisig? AI can't know this.

Source