
Keep the comparison honest: published findings and exploit incidents are different datasets. They cannot tell us how many attacks audits prevented. Use the talk to question coverage assumptions and plan the security work around the audit.
✓Book. Secure. Relax. Oak Security offers audits, penetration testing, training, and advisory.

Keep the comparison honest: published findings and exploit incidents are different datasets. They cannot tell us how many attacks audits prevented. Use the talk to question coverage assumptions and plan the security work around the audit.

Take one path to funds. Trace it through the contracts, signer devices, approval process and recovery procedure. Mark what your planned review covers. Give each uncovered part an owner and a review method.

Where does your audit scope end? In “The Audit Gap,” @christianvari_ compares public audit findings with real exploit losses, bringing key custody and operations into the coverage discussion. Watch the @summit_defi talk in under 5 minutes: youtube.com/watch?v=96YwISrz…

A model flags a bug. What would disprove it? @christianvari_ walks through Oak's adversarial AI audit workflow in “Don't Trust the First Model.” Separate discovery, challenge the evidence, keep human judgement. Watch the @summit_defi recording: youtube.com/watch?v=7otm-JgM…

Planning an audit? Tell us what you're building, what is changing and which assumptions worry you. Talk to Oak about human-led, AI-assisted audits and how TRACE can help define the scope. oaksecurity.io/contact

At Oak, human researchers agree the scope, manually audit the code and verify findings. Hive supports discovery and adversarial validation. Reach out to talk through your next release.

Add privileged roles, external dependencies and the assumptions you want challenged. Mark anything that changed since the last review.

Preparing for a human-led, AI-assisted audit? Start with the exact revision and the behaviour your system must preserve.

A compromised account can become a malicious build, then a production change. TRACE follows authority across protocol rules, systems and people. For each boundary, ask what still holds if the source is compromised. Methodology: github.com/oak-security/TRAC…

20 of 257 incidents account for 66.3% of recorded losses in our 2022-Q3 2026 study. For your threat model: which failure could cause a loss you cannot absorb, and what limits the damage? Report: research.oaksecurity.io/ Incident data: @RektHQ.

Two models agreeing can still mean two models sharing a blind spot. Hive keeps discovery separate and records the reasons behind validation. A senior researcher reviews the code and the evidence before accepting a finding.

TRACE makes that review explicit. AiTM applies the workflow with AI assistance; we're preparing its open-source release. The researcher still needs to judge whether the analysis holds.

Then inspect the path. What must the attacker control? Which invariant would fail? Which step is assumed rather than established?

What should you ask of an AI-assisted threat model? Start with the evidence: which source supports this actor, permission or trust boundary?

An architecture diagram shows connections. A threat model also needs to show authority. Who can change the build, approve a transfer or recover an account? Follow those paths across protocol rules, infrastructure and people. That is where TRACE starts.

We're preparing to open-source AiTM, our AI Threat Modeller, built on TRACE. It uses AI to turn specs and code into a structured threat model for human review. Release details will follow. Register your interest: oaksecurity.io/ai-threat-mod…