Twelve weeks of live fire. Zero breaches.
A real HYVE Ether OS beta key has sat behind Hyve Raptor for 93 days, in the open, with the public API key published. This is the honest running review — every number below is a live aggregate from the range’s own database, with our commissioning tests and self-run red-team broken out separately.
The traffic, honestly
Not every request is an external attacker, and we won’t pretend otherwise. Here is where the 33,792 logged requests actually came from:
Synthetic load we fired on day one to prove the ingest + classify pipeline holds under a flood. Not external actors.
Real internet actors — 3,176 unique sources across 79 countries. 4,676 carried a genuine attack signature.
An independent security audit and a self-run nation-state red team, including a direct database assault with the public key. All logged, all repelled.
How they attacked
The real-world picture of a public honeypot: overwhelmingly opportunistic recon and automation — admin/login probing, credential-file hunting (.env / .git / .aws), WordPress takeover attempts, and volumetric noise — not a targeted 0-day campaign. The exotic exploit families (Log4Shell, SQLi floods, template injection) showed up mostly in our commissioning and red-team runs; Raptor holds a signature for each regardless. Just as telling as what it blocks is what it doesn’t: legitimate visitors browsing the range are scored clean and let straight through.
We will not claim a perfect record. A full adversarial audit of this classifier on 2026-09-07 found a real false-positive class, so here it is. A field separator inside the detection engine supplied a newline that the command-injection rule matched on — so any client whose User-Agent began "curl" or "wget" was scored as command injection, rated high, and auto-banned. 27 of the 126 command-injection labels in this corpus are wrong because of it. Fixed the same day and pinned by regression tests that fail against the old code. The affected rows are left exactly as they were recorded rather than quietly relabelled — a range that edits its own history is not evidence.
Where they came from
Last 31 days at a glance
The day-one spike is the resilience stress test. After commissioning, organic traffic settled into a steady baseline of opportunistic scanning and bot noise — on the order of a hundred-plus requests on a typical day — punctuated by the audit and red-team runs.
Severity mix
Every one of the 1,471 high- and critical-severity events was classified and repelled on contact — 7,295 attacks defeated all-time, 0 breaches, and 902 vault codes submitted that never matched. 93 days in the open, public key and all, and the number that matters most has not moved off zero.
What we've shipped
A 30-finding external audit of the range's own code. Every finding closed: a transactional single-winner prize path, cross-instance rate-limiting, consent-gated telemetry, single-owner admin authorization (MFA-ready), plus lint/tests/CI gates.
We attacked our own range: a direct database assault with the public key (RLS returned nothing), then advanced WAF-evasion, CVE chains, injection and AI-poisoning. Ejected on first contact. The vault never moved.
New signatures for AI-attack tradecraft (prompt injection, jailbreaks, Pliny markers, glitch tokens, invisible-Unicode smuggling), credential-file recon (.env / .git / .aws), and the Vite dev-server file-read family — refreshed on a weekly cadence as fresh tradecraft appears on the range, ejecting on first contact.
Everything the range learned was ported into HYVE Overlord and re-published, so a customer's install detects the same tradecraft the live range does.
The vault
902 people have submitted a code to the vault. None matched.The code lives in a table only the server can read — even with the public key, the front end can’t reach it. That’s why 30 days in the open changed nothing.
Figures are live aggregates from the range database, refreshed periodically. Commissioning stress-test and our own red-team traffic are broken out and excluded from “organic” totals.