Matthew Bowman

Reconnaissance output is a flat text dump, and the public internet maps are frozen archives that each render one layer of the whole internet — neither gives you a live, navigable view of a single target's complete footprint, or of the shape and timing of its attack surface.

An async, latency-tiered OSINT engine streams every probe result the millisecond it returns over a WebSocket to a 3D force-graph: domains, subdomains, IPs, prefixes and autonomous systems render as stars, planets, moons and suns, with BGP and DNS relationships drawn as gravitational lanes. Everything is keyless, and passive by default (DNS, Certificate Transparency, BGP whois, local GeoIP), and each node ignites the instant it arrives — so probe latency becomes the choreography rather than a loading bar. A 'home base' mode turns the same engine inward, mapping your own host outward in concentric shells and flagging live egress that falls outside your normal network neighborhood.

Turns recon from a static list into a live, explorable map where an attack surface's topology and timing are legible at a glance — and, pointed inward, into a defensive instrument that surfaces anomalous egress by construction.

New subdomains, certs, and look-alike infrastructure appear constantly — phishing and shadow assets hide in the volume.

Continuously ingest public CT logs, extract and normalize domains, correlate against tracked roots, and surface only the new or anomalous as actionable intel.

Early warning on phishing infrastructure, subdomain sprawl, and shadow assets — attack-surface monitoring that runs unattended.

All AI evaluation tools score a model's vulnerability, but none measure whether a defense middleware actually helps or by how much — you get a baseline and a prayer.

Point at any Ollama-hosted model, run all four engines in sequence (deterministic seed probes → NVIDIA garak probes → inline defense re-scoring → delta-driven report), then compare baseline vs. defended scores as a first-class CLI primitive: 'assay delta baseline defended.' Ships a premium HTML report and a multi-run dashboard, so the whole evaluate → delta → dashboard pipeline is reproducible and auditable end to end.

Turns 'is it secure?' from vibes to a letter grade, and 'does the defense help?' from guesswork to a measured percentage-point lift. The delta wedge makes it the only honest defense-evaluation tool in the OSS AI-security space.

Automated decision systems optimize for being right and forget to optimize for surviving being wrong — a single bad sizing call ends the game.

A research-to-decision pipeline behind a layered risk gate, paper-trade execution, and a live operations dashboard. Most candidates are rejected by design; the system acts only when conviction and risk both clear.

Capital-preservation-first automation: it does nothing unless conviction and risk both clear — 'no decision' is the default, not a failure.

Every permissionless blockchain suffers MEV/front-running. Proposed defenses claim extraction resistance, but none are systematically tested under adversarial stress. The space has no framework for auditing a mechanism's boundary conditions before deployment.

Rigorous iterated adversarial mechanism design: propose a hypothesis, simulate it (Python stdlib-only, deterministic and reproducible), subject it to adversarial review, then falsify or refine it. The output is a set of formal impossibility results, a catalog of dead ends, an audit checklist for any extraction-resistance claim, and an honest shippable spec built on existing batch-auction and threshold-encryption protocols.

The constraint framework is the product — a general design methodology for any protocol claiming extraction-resistant sequencing. Turns 'is it MEV-resistant?' from marketing copy into a falsifiable audit. A monetary-base extension applies the same safety principle as a minting rule for an engine-backed currency, where the impossibility does not bind.

Most organizations adopt AI systems without understanding their security surface. Incumbent security firms treat AI risks as a checkbox — generic penetration tests miss agent-specific threat models, injection surfaces, and supply-chain vectors.

Advisory-altitude assessments that go deeper than a checklist: architecture reviews with threat models drawn from production AI stacks, adversarial red-team evaluations using custom injection-evaluation harnesses, and prompt-injection defense designs informed by first-principles research. Each engagement delivers a concrete findings catalog with ranked mitigations, not a scorecard. The firm's own R&D platform (Meridian, Division, Sentinel, Seal) provides direct operational insight into how autonomous adversaries think and move.

Turns AI security from a checkbox exercise into a defensible architecture — organizations understand their real attack surface, not the one a generic checklist covers. The consultancy itself is a working proof that the practitioner's own research pipeline is the strongest signal of true expertise.