Adversarial AI evaluation
Red-team model behavior, agentic workflows, retrieval paths, and tool permissions before they reach sensitive systems.
Applied AI + Information Security
Aesoteric
Applied research for AI systems that need to survive real adversaries, ambiguous data, and production constraints.
Applied AI research
Information security
Adversarial evaluation
Production hardening
Research operating loop
Move from unclear risk to tested evidence.
Aesoteric works where AI capability, security engineering, and product reality overlap. The output is not theater. It is evidence engineers can reproduce, fixes they can ship, and monitoring that keeps learning after launch.
eval-runner.ts
ThinkingGreppingReadingEditingDone
threat_model.scope()
- tools: browser, files, identity
- secrets: production, customer
- exits: approval, audit, revoke
findings.push({
path: "retrieval -> tool -> data export",
severity: "high",
fix: "bind tool grants to verified intent"
})
Capabilities
Research depth with implementation pressure.
Secure model integration
Design architectures that isolate secrets, bound actions, preserve auditability, and degrade cleanly under attack.
Agentic system hardening
Map tool calls, memory, identity, and approval surfaces so autonomous software stays inside real operational limits.
Detection research
Turn model traces, security telemetry, and product events into signals teams can investigate and act on.
Data perimeter design
Keep confidential data, regulated context, and internal reasoning artifacts separated by policy and by implementation.
Technical diligence
Assess AI products, vendors, and security claims with concise findings, implementation risk, and remediation paths.
Engagements
Small teams, direct work, concrete outcomes.
Aesoteric is built for high-trust technical work: fewer layers, tighter feedback, and findings that survive review by engineering, security, legal, and leadership.
Security review
A focused assessment of an AI feature, agent, workflow, or model integration with prioritized findings.
Research sprint
A short applied research cycle for a hard question: feasibility, threat model, prototype, and decision memo.
Embedded build
Hands-on collaboration with engineering and security teams from architecture through production launch.
Evidence package
Useful artifacts, not vague assurance.
Every engagement closes with enough technical detail to reproduce the issue, prioritize the work, and verify the fix.
- Threat models grounded in real system behavior
- Attack paths with reproducible evidence
- Architecture changes engineers can implement
- Detection and monitoring recommendations
Bring the hard AI security question.
Share the system, risk, or research problem. Aesoteric will help turn it into a plan that can be tested.