Security infrastructure for agentic AI

Give every AI agent an identity, a conscience, and a memory.

ARTONexa secures autonomous agents end to end: a forgery-proof identity layer, a runtime firewall that stops harmful actions mid-trajectory, and a tamper-evident ledger that proves what happened.

Request early access View on GitHub ›

Three tools, one platform — Tessera · VerityFlux · Vestigia. Source-available under BSL 1.1.

console · triage

Correlated alerts · session sess_9f3c2a71● live

CRITICAL Tool executed 12s after A2A reasoning alertreasoning_a2a_then_execution · payments-reconciler-07 87

HIGH Token issued 40s after threat detectionrapid_token_after_threat · onboarding-bot-11 71

MEDIUM Cross-agent memory-poisoning signalcross_agent_memory_then_execution · summarizer-07 52

The governance gap

Most organizations can't answer the basic questions.

Who is this agent, what is it authorized to do, and what did it actually do? ARTONexa exists to make those answerable.

01 // IDENTITY

No Identity

Agents run under shared service accounts and inherited credentials provisioned for something else. You cannot govern what you cannot identify.

02 // AUTHORIZATION

No Authorization

Role-level grants can't express action-level constraints. An agent with read access to a database has read access to the entire ecosystem.

03 // AUDIT

No Audit Trail

Logs exist; audit-ready evidence does not. A log that can be modified is not evidence of what happened — only of what someone claims happened.

04 // ACCOUNTABILITY

No Accountability

When an agent causes systemic harm, accountability is distributed so broadly it belongs to nobody. Named, revocable responsibility is absent.

The platform

One platform, from identity to evidence.

Most agent security stops at a prompt filter. ARTONexa covers the whole lifecycle of an agent action — who it is, what it's doing, and what it did.

01 · IDENTITY & ACCESS

Tessera

Zero-trust IAM for autonomous agents: every agent gets a unique scoped credential with DPoP token binding, delegation-chain narrowing across five depth levels, and Redis-backed revocation that propagates immediately.

DPoP BindingDepth-5 DelegationInstant Revocation

Repository source ›

02 · RUNTIME DEFENSE

VerityFlux

Session-level behavioral enforcement: 27 adversarial detectors and trajectory scoring on the complete conversation. Multi-turn attacks that per-request parsing catches at ~0% are flagged before critical in ~44% of runs on a realistic corpus. LLM-based scoring runs on NVIDIA NIM.

27 DetectorsTrajectory Engine~0% → ~44% Multi-TurnNVIDIA NIM

Repository source ›

03 · EVIDENCE & AUDIT

Vestigia

Tamper-evident forensics: SHA-256 hash-chained logging at the interaction and tool-execution level, OpenTelemetry distributed tracing, secure SIEM forwarding, and automated PII sanitization.

Hash-Chained LogsOpenTelemetrySIEM Pipeline

Repository source ›

The difference

A closed accountability loop.

The three tools share one session thread, so an incident flows from identity to enforcement to evidence to containment — without a human stitching it back together.

01 · TESSERA

Identity

Every agent gets a scoped, revocable credential before it acts.

02 · VERITYFLUX

Enforcement

Runtime scoring stops a harmful trajectory mid-session.

03 · VESTIGIA

Evidence

The decision and its context are hash-chained into the ledger.

04 · TESSERA

Containment

Revoke the credential — and the action stays attributable, for good.

One console

See the whole incident in one place.

→
One session thread ties identity, runtime, and audit events into a single timeline. No copying IDs between three dashboards.
→
Cross-component anomaly detection flags patterns no single tool sees. Like a tool firing seconds after an alert.
→
Two clicks to contain: revoke the credential, export the evidence. The loop closes where it started.

Explore the source ›

Source-available · BSL 1.1 · Python

Clone it. Run it in under 10 minutes.

The core framework is source-available — inspectable and runnable, with no framework lock-in.

Tessera

Python · FastAPI · Redis · DPoP

Zero-trust IAM for AI agents. Clone and run in under 10 minutes.

github.com/ArksherX ›

VerityFlux

Python · 27 Detectors · NVIDIA NIM

Plugs into any LLM agent pipeline. No framework lock-in.

github.com/ArksherX ›

Vestigia

Python · OpenTelemetry · PostgreSQL · Docker

14 orchestrated services. Production-ready deployment out of the box.

github.com/ArksherX ›

Run Tessera locally — under 10 minutes

$ git clone https://github.com/ArksherX/ARTO
$ cd ARTO/tessera
$ docker compose up          # postgres · redis · api · dashboard

Contributions welcome via GitHub Issues and Pull Requests · Community: Discord

Empirical benchmarks

Single-turn monitoring is structurally blind.

Trajectory monitoring recovers a measurable part of what per-request parsing misses.

~0%

Per-request monitoring catch rate

~44%
Multi-turn attacks caught before critical (natural-language corpus)

76.9%

Upper bound on a keyword-scored synthetic benchmark

~8%

False-positive rate

The Accumulation Problem — Single-turn monitoring is structurally blind to multi-turn attacks; trajectory monitoring recovers a measurable part of what it misses. The ~44% figure is measured on a realistic natural-language corpus (~8% false-positive rate); 76.9% is an upper bound on a keyword-scored synthetic benchmark.

Read the full evaluation report →

Controlled agency

Weekly research on agentic AI security and governance.

An ongoing weekly series on the monitoring gap, the enforcement gap, MCP trust architecture, agent identity, the governance specification, the memory attack surface, and the future of AI defense. Published every Tuesday.

22The Gap Between Observation and Control→ 21The Control Nobody Called→ 20The Maintainer Said No→ 19You Approved a Filename, It Wrote Your SSH Keys→ 18Nobody's Job→ 17Can You Prove Your Log Wasn't Edited?→ 16Authority You Can Argue With Isn't Authority→ 15The Competence Illusion→ 14The Future of AI Defense→ 13What I Got Wrong About AI Security→ 12The Blank Spaces on the AI Safety Map→ 11The Confidence Paradox→ 10The Hidden Cost of AI→ 09The Experience Hijack→ 08The Memory Problem→ 07The Governance Specification→ 06The Guidance Gap→ 05The Protocol That Trusts Everything→ 04The Enforcement Gap→ 03The Identity Vacuum→ 02The Trajectory Problem→ 01AI Autonomy and the Future of Work→

Peer recognition & contribution

Recognized across the AI-security community.

NVIDIA Inception

Member · ARTONexa Ltd.

NVIDIA garak

Contributor · PR #2272

Black Hat India

Accepted Speaker

google/CAGE

Merged Contributor · PR #241

OWASP WrongSecrets

Merged Contributor · PR #2724

DEF CON 33

Policy Track Speaker

Black Hat MEA 2025

Research Presenter

OWASP AI Exchange

Core Framework Author

AIVSS v0.8

Named Specifications Reviewer

OASec 2026

Founding Member Circle

UN Global Dialogue

AI Governance Contributor

Swift Centre

Published Policy Contributor

Contributor

Miracle Owolabi

Founder & AI Security Researcher, ARTONexa Ltd.

ARTONexa Ltd. was founded from years of empirical adversarial research on agentic AI systems. The ARTO Security Framework (Tessera, VerityFlux, and Vestigia) was built to operationalize those findings and close the governance gap that every major AI security framework now specifies as a foundational requirement.

Miracle is a core author of the OWASP AI Exchange, a named reviewer on the AI Vulnerability Scoring System v0.8, and a speaker at DEF CON 33 and Black Hat MEA 2025. He contributes to international AI governance through the UN Global Dialogue on AI Governance, the Swift Centre Bridge the Gap policy platform, and ISO/IEC standards workgroups.

LinkedIn → Newsletter → Writing → Email →

Enterprise integration

Work with ARTONexa

The core framework is source-available under BSL 1.1. Production-grade deployments, managed cloud environments, implementation engineering, and continuous compliance mapping are available for enterprise networks.

GitHub Newsletter LinkedIn Discord

Secure your agents before they act.

ARTONexa is onboarding early design partners. If you're deploying autonomous agents and need identity, runtime defense, and audit that work as one, let's talk.

Request early access Read the research

Or email info@artonexa.com