Research / AI Usage at Oak Security
Meet Hive.
Led by humans.
Challenged by AI.
Hive is Oak Security’s continuously evolving proprietary multi-model AI harness. It brings independent model perspectives, adversarial validation, and technology-specific threat corpora into our security researchers’ workflow.
Our commitment in every audit
Every line reviewed.
Every finding verified by a senior researcher.
We use Hive in all audits. At least one senior human security researcher always performs a manual audit of every line of in-scope code, then verifies all findings before they enter the final report. Human expertise and accountability remain central throughout.
The methodology
Independent discovery.
Adversarial scrutiny.
A plausible explanation is only a starting point. Our human-led, multi-model adversarial methodology tests candidate vulnerabilities against the actual code and the system’s assumptions before a researcher decides whether they belong in a report.
-
Scope: establish the target and coverage
Human researchers agree the target revision, scope, and security priorities, then map components, entry points, privileged operations, and invariants. Hive records which in-scope files each model reads, retries missing work where feasible, and exposes remaining gaps. File access alone does not establish understanding or safety.
-
Discover: separate blind and informed passes
Multiple models examine the code without sharing candidate findings or discovery reasoning. Each model uses separate contexts for a blind pass, which withholds the specialist vulnerability catalogue, and an informed pass, which supplies stack-specific threat knowledge. Each candidate records its detecting model and pass, code location, attacker, preconditions, and potential harm.
-
Deduplicate: merge by root cause and correction
Candidates are merged when they describe the same underlying defect addressed by the same targeted correction. Shared locations or similar titles alone are insufficient. Detection history is preserved so independent rediscovery remains visible to the reviewer.
-
Validate: try to disprove every claim
Every configured model challenges every surviving finding, including its own, in a fresh context without discovery reasoning or other validators’ votes. Validators inspect callers, guards, and reachable attack paths against the unmodified source. A timeout or missing verdict is recorded as missing evidence; it cannot establish or refute a finding.
-
Review: make an accountable human decision
At least one senior human security researcher manually audits every line of in-scope code, then verifies all findings. The reviewer receives the scope, coverage, run limitations, detection history, and validators’ reasons and disagreements. Human researchers decide validity, severity, impact, and what reaches the client report.
Independence here means keeping discovery work separate. Models can still share blind spots. Confidence is a signal for human review, distinct from severity and from a measured probability that a finding is correct. Reasoned validation, a demonstrated proof of concept, and human acceptance are separate evidence and decision states.
Threat corpora
Security knowledge for the stack in front of us.
Our threat corpora organise vulnerability classes and detection guidance for several technology stacks. They inform the informed pass while the blind pass leaves room for weaknesses that have not been catalogued. Blind refers to withholding our specialist catalogue; models retain their prior knowledge and biases. We evolve the corpora alongside the harness as our research and the technology change.
Smart contracts & protocols
Solidity, CosmWasm, Cosmos SDK, and Noir: security knowledge tailored to the execution environment and its trust boundaries.
Applications & infrastructure
Rust, Python, C/C++ memory safety, cloud infrastructure and infrastructure as code, alongside general software vulnerability classes.
Economic & cross-cutting threats
DeFi economic attacks and general weaknesses such as access-control failures, injection, cryptographic misuse, and state-machine flaws.
Why this approach
Less noise.
Fewer false positives.
We have tested this methodology against many other approaches. In our internal comparisons, its key advantage is substantially less noise and fewer false positives. Independent discovery, adversarial checks, and senior human verification help us focus on findings that stand up to scrutiny.
We keep refining Hive as models, threats, and codebases evolve. These observations reflect our internal testing; results depend on the codebase, scope, and model configuration.