CYBERSECURITY × AI
│
├── AI / Agent Security
│   ├── Prompt injection & tool abuse
│   ├── Agent identity & authorization
│   ├── Capability security & provenance
│   └── MCP security
│
├── Security Infrastructure
│   ├── Reference monitors
│   ├── Vulnerability scanning
│   ├── EDR / XDR for agents
│   ├── Security observability
│   └── Incident analysis
│
├── AI Reliability
│   ├── Failure attribution
│   ├── Chaos / resilience testing
│   ├── Evaluation systems
│   └── Runtime safeguards
│
└── AI Systems
    ├── LLMs & SLMs
    ├── Interpretability
    └── Local inference

Security architecture built around prevention before detection.

• Dual-LLM isolation with privileged and quarantined contexts
• Reference-monitor enforcement before tool execution
• Capability policies, provenance tracking, and least privilege
• Static taint analysis for agent/tool integrations
• Session replay, attack DAGs, and incident reports
• Adaptive attack harness with obfuscation and MCP-specific tests
• AgentDojo + AgentDyn evaluation

Authorization infrastructure for systems where agents delegate work to other agents.

• Invocation-Bound Capability Tokens
• Chained Attenuation Tokens
• Scope narrowing across delegation hops
• Offline verification and revocation
• SPIFFE/SPIRE workload identity
• MCP OAuth 2.1 + PKCE
• SLSA-style provenance
• W3C Verifiable Credentials + A2A

Delegation may narrow authority — never expand it.

Infrastructure for answering: what is an agent built from, and what is it allowed to do?

• SPDX 3.0 AI Profile + CycloneDX ML-BOM v1.7
• Agent discovery and inventory
• Configuration drift detection
• OPA policy enforcement
• Identity / tool / data / model risk scoring
• EU AI Act, NIST AI RMF, and ISO 42001 crosswalks

Treats agent failures as causal systems problems, not just log messages.

• OpenTelemetry-native instrumentation
• Causal graph reconstruction
• Counterfactual failure attribution
• MAST+ failure classification
• Contextual-bandit recovery selection
• Uncertainty-calibrated human escalation
• Reliability scoring and topology risk

Open-source AI/ML auditing platform combining security assessment, red teaming, interpretability, alignment checks, and unified risk scoring.

• Vulnerability scanning
• Red-team attack simulation
• Behavioral probes
• Mechanistic interpretability
• Alignment auditing
• 7-dimension risk scoring
• SIEM integrations
• CLI + REST API + Docker

Security tooling for MCP implementations: code vulnerabilities, dependency CVEs, hardcoded secrets, MCP flaws, and tool poisoning.

Interpretability infrastructure for activation harvesting, SAEs, cross-layer transcoders, attribution graphs, feature interpretation, steering, and drift monitoring.

Self-hosted monitoring for public attack-surface exposure, open ports, risk scoring, and breach signals.

Interactive 3D exploration of quantum cryptography and attacks using Qiskit + Godot, including BB84 and Grover's algorithm.

Most of my recent work follows the same loop:

       AI SYSTEM
           │
           ▼
      ┌──────────┐
      │ OBSERVE  │
      └────┬─────┘
           ▼
      ┌──────────┐
      │  ATTACK  │
      └────┬─────┘
           ▼
      ┌──────────┐
      │ MEASURE  │
      └────┬─────┘
           ▼
      ┌──────────┐
      │ ENFORCE  │
      └────┬─────┘
           ▼
      ┌──────────┐
      │  IMPROVE │
      └──────────┘

I'm especially interested in the layer around the model: identity, permissions, tool execution, telemetry, evaluation, failure handling, and security controls.

• 🔐 Security architectures for AI agents
• 🧩 MCP and tool-use security
• 🪪 Agent identity, authorization, and delegation
• 🛡️ AI vulnerability assessment and red teaming
• 📡 AI observability and OpenTelemetry
• 🧪 Agent evaluation and reliability engineering
• 🧠 Mechanistic interpretability
• 💻 Local LLM inference and small-model experimentation

Languages — Python · TypeScript · SQL · Bash

AI / ML — PyTorch · Hugging Face · Transformers · scikit-learn · Qiskit

AI Systems — LLMs · SLMs · Agents · MCP · OpenTelemetry · vLLM · llama.cpp

Security — Static Analysis · Taint Analysis · Red Teaming · IAM · OAuth · mTLS · Threat Modeling

Infrastructure — FastAPI · PostgreSQL · Redis/Valkey · Docker · Kubernetes · Terraform · GitHub Actions

I build from first principles — from neural networks and LLM internals to the infrastructure required to deploy, observe, evaluate, and secure agentic systems.

Some projects are experimental. Some are research-adjacent. Some are production-oriented. I prefer making the work inspectable over pretending everything is finished.