Severity Scoring for AI Vulnerabilities, v1.0
How LogicLeak rates AI security findings against CVSS, OWASP LLM Top 10, and engagement-specific impact criteria. Includes scoring rubrics and worked examples from sanitized engagement findings.
RESEARCH · METHODOLOGY← BACK TO RESEARCH
LogicLeak Methodology
Our methodology is open. We publish what we believe about adversarial AI security so clients can challenge it and the field can build on it.
Methodology is the slower-moving stream in LogicLeak Research — published frameworks, scoring systems, and operational philosophies that govern how we run engagements and publish findings. Methodology documents are versioned and revised as we learn. The first document publishes at firm launch; the others follow on a quarterly cadence.
// FIRST PUBLICATION · LAUNCH
Published at firm launchFRAMEWORK · V1.020–30 pages · Living document
engagementadversarialoperational
The Adversarial Probing Methodology, v1.0
Our 12-stage process for red-teaming production AI systems, from reconnaissance through remediation handoff. This document describes what we do, why we do it, what we don't do, and how each phase produces evidence the next phase relies on. Updates published quarterly as the methodology evolves.
// PUBLISHED METHODOLOGIES
Current versioned documentsFRAMEWORK · V1.0ACTIVE
The Adversarial Probing Methodology
Our 12-stage red-teaming process from reconnaissance to remediation handoff.
Additional methodologies publish on a quarterly cadence — see Upcoming below.
// UPCOMING METHODOLOGIES
In drafting; publication dates approximateSanitization & Disclosure Policy, v1.0
How LogicLeak strips client identification from published research while preserving technical fidelity. Includes sanitization protocols, consent workflows, and review processes.
Multi-Agent Threat Modeling, v1.0
Our approach to threat modeling AI systems with three or more communicating agents. Covers trust boundary identification, escalation path analysis, and graph-aware attack surface mapping.
Rules of Engagement, Published Version
A formal version of our operating commitments, published as a versioned methodology document rather than only existing on /firm/rules-of-engagement. Versions track over time.
// VERSIONING APPROACH
How methodology documents evolve over time01
Methodologies are versioned, not edited silently.
When a methodology evolves, we publish a new major or minor version with a change log. Previous versions remain accessible. The current version is always marked ACTIVE; superseded versions are marked accordingly.
02
Major versions reflect structural change.
A major version bump (v1 → v2) means a structural change in methodology — phases added or removed, scoring criteria fundamentally changed, scope expanded or contracted. Major versions are accompanied by a published rationale for the change.
03
Minor versions reflect refinement.
Minor versions (v1.0 → v1.1) reflect refinements, clarifications, or additions that don't change the core methodology. Minor version changes are documented in the change log but don't require separate rationale.
04
Versions reference the engagements that drove change.
Where a methodology revision is driven by lessons from specific sanitized engagements, the change log references them. This makes our learning visible — the methodology is not handed down from theory, it's revised against reality.
05
All versions remain accessible.
We do not delete previous methodology versions when new ones publish. Each historical version remains accessible at a permanent URL so research that referenced an older version remains coherent.
Methodology stream publishes quarterly. First document live; next publications Q3 2026.
SEE OTHER RESEARCH STREAMS →