When the Autonomous Attacker Meets an Invisible Network

On July 29, 2026, the team behind Darkmoon — an open-source, fully autonomous AI penetration-testing platform built by ASC-IT in Toulouse, France — pointed their system at OpenNHP's public demo environment and published the results.

We want to start with the only thing that actually matters here: thank you. Darkmoon ran a real, reproducible, non-destructive assessment against a live target, and then published the full report — methodology, tooling, findings, and all — in the open. That is precisely the kind of transparent, independent scrutiny we asked for when we invited the security community to try to break OpenNHP at IETF 126. Darkmoon did exactly that, rigorously, and told the world what happened either way.

Darkmoon isn't a person running nmap by hand. It's an LLM orchestrator directing specialist sub-agents through reconnaissance, discovery, exploitation, and reporting — fully autonomously, at machine speed. That is exactly the class of adversary we built OpenNHP for. Offensive AI tooling is not a hypothetical future threat; it's an open-source project on GitHub today, and it's only going to get faster and more capable.

If autonomous AI is about to make reconnaissance and exploitation effectively free, then testing a network-hiding protocol against exactly that kind of adversary isn't optional — it's the whole point.

The assessment followed NIST SP 800-115 and OWASP testing guidelines and combined multiple independent reconnaissance techniques against OpenNHP's demo hosts:

• HTTP/HTTPS requests via curl
• Port scanning with naabu across common ports (22, 80, 443, 8080)
• TCP connect probes via bash /dev/tcp
• ICMP probing

Four independent methods, aimed at three NHP-protected hosts (server2, ac2, and the Access Controller) sitting behind OpenNHP's Network-infrastructure Hiding Protocol (NHP).

It's important to be precise about that second number, because we don't want to bury it. Darkmoon's 51 findings — 4 high, 10 medium, 17 low, 20 informational — were against the public auth backend, a component we intentionally expose on our demo site as a teaching surface so visitors can see what a normal, unprotected web service looks like by comparison. The high-severity findings involved cross-origin credential exposure in that same intentionally public component. That surface was never behind NHP, and finding real issues there is Darkmoon's tooling working exactly as intended.

On the hosts actually protected by NHP, Darkmoon's scanners didn't find a misconfiguration, an interesting banner, or a soft edge to poke at. They found nothing to look at in the first place.

Darkmoon's own conclusion is the one we'd choose ourselves: OpenNHP doesn't make services unbreakable, it makes them undiscoverable until authenticated. As they put it, the approach "relocates the entire attack surface onto credential secrecy" by removing the one precondition every attack chain depends on — discoverability. Their report frames it well: authentication happens before connection, not after.

That is the network-hiding thesis, tested by an autonomous adversary instead of asserted by us. A clean result against a real, modern, AI-driven attacker is meaningful evidence that the approach holds up — not proof that it always will, but a real data point in its favor, produced independently and published in full.

Zero Trust has spent a decade teaching the industry not to trust a connection just because it came from inside the perimeter. Darkmoon's results point at the next step in that evolution: don't just refuse to trust an unauthenticated connection — refuse to let one exist in the first place. Verify identity before revealing that there's anything there to connect to at all.

We'd call that Zero Visibility, and we think it's where Zero Trust architecture is headed. Autonomous AI attackers are the reason the industry needs to get there faster.

Security should never depend solely on the claims of its creators. OpenNHP is an open-source project because security improves through transparency, peer review, and independent validation.

The Darkmoon evaluation demonstrates exactly the type of collaboration we hope to see more of:

• Independent offensive testing
• Public methodologies
• Reproducible experiments
• Constructive technical discussion

We invite researchers, penetration testers, universities, and AI security teams to evaluate OpenNHP using their own tools and methodologies. Whether your platform is based on autonomous AI agents, traditional penetration testing, or entirely new offensive techniques, we welcome independent assessment.

One clean report from one excellent open-source tool is not the end of this conversation — it's an invitation to more of it. If something gets through, we want to know immediately, and we'll say so publicly, the same way Darkmoon did. The live demo stays online specifically for this reason. Come try.

Artificial intelligence is changing both sides of cybersecurity. Attackers are becoming autonomous. Defenders must become proactive.

OpenNHP represents one possible evolution of Zero Trust — from simply "Never Trust, Always Verify" toward an additional principle increasingly relevant in the AI era:

We sincerely thank the Darkmoon team for contributing to this discussion. Their evaluation helps demonstrate an important point: when autonomous attackers become faster, smarter, and cheaper, making infrastructure invisible to unauthorized users becomes one of the most effective ways to prevent attacks before they begin.

Thank you again to the Darkmoon and ASC-IT team. This is exactly what independent, adversarial, published testing should look like — and the OpenNHP project is stronger for it.