Sherlock has publicly launched Sherlock Audit Engine, revealing a security auditing platform the company had largely kept under wraps while testing the model with protocol teams.
Audit Engine operates one layer above individual AI auditors, coordinating several approaches to vulnerability discovery inside the same review.
Frontier LLMs, purpose-built AI auditors and AI-enabled security researchers work against the same codebase and context. Sherlock handles orchestration across the engagement, with findings judged, validated and deduplicated before being consolidated into one final audit result.
That model also sheds new light on one of Sherlock’s more unusual engagements this year.
Was Polygon an Early Look at Audit Engine?
In June, Sherlock put Polygon’s Heimdall V2 through a review involving a broad field of AI auditing systems and security researchers.
Heimdall V2 is the consensus client at the core of Polygon PoS, making it a consequential codebase for an early deployment of the model.
With Audit Engine now public, the engagement appears to have served as a proving ground for the platform Sherlock is bringing to market.
Automated AI auditing systems emerged among the strongest performers for overall coverage, while different systems and researchers surfaced different portions of the overall issue set. The result reinforced a central idea behind Audit Engine: no single approach captured the full security picture.
The platform measures those differences directly. Teams can see which systems delivered broad coverage, which maintained high precision and where different approaches contributed complementary security signal.
That becomes increasingly relevant as the underlying technology changes.
In July, Google DeepMind introduced Gemini 3.5 Flash Cyber, a cybersecurity-specific model designed to find, validate and patch vulnerabilities quickly, another indication of how fast specialized AI security capabilities are developing.
For security teams, the best available mix is therefore a moving target.
Audit Engine is designed to incorporate new models, auditors and researcher methodologies as they emerge, while giving protocols a consistent environment for measuring what actually performs well against their code.
Bitrue