Breadth at low cost: one accepted finding from a $3 scan

The newest completed scan is the broadest one run so far on this target, and the cheapest per hypothesis. 101 attack hypotheses proposed, 54 actually read against the source after a cheap pre-screen rejected the rest, 705 model calls, 17.0M tokens, $3.01 total spend, 308 minutes. Thirty were refuted with a cited defense, twelve came back unverifiable-context, ten unverifiable, one candidate — and one survived the engine's adversarial refutation pass.

The funnel

Stage	Count
Hypotheses proposed	101
Pre-screen rejected	47
Analyzed	54
Refuted	30
Unverifiable: context	12
Unverifiable	10
Candidate	1
Confirmed	1

The hunter model that generated hypotheses was the cheap, fast tier; the triager that ran bulk verification was a mid-cost model. Neither is the most expensive model available to the engine. The breadth-at-low-cost thesis this engine has been built around — cheap models for the wide search, stronger models reserved for the few calls where judgment actually changes the outcome — is what produced this funnel.

The finding that survived

One HIGH-severity, missing-authentication finding survived the engine's adversarial refutation pass: a control path in a network-facing worker accepts a privileged command from a connection that was never required to authenticate, because the authentication gate that exists in the file does not cover that particular command's dispatch. It is a simple mechanism, cited to specific lines, and a human could check the claim in minutes by reading the two functions involved.

We are deliberately not publishing the file, the command, or the exact mechanism here. The target has a real fix to ship first, and this project's standing rule — stated plainly on this site — is that hands-on reproduction and responsible disclosure come before anything is written up as proven, and the target itself stays unnamed until a fix is out. The same discipline applies to exploit detail: a $3 engine run producing a plausible, well-cited lead is a result worth reporting; handing the internet a working bypass for an unpatched, open-source service is not. The write-up with the reproduction steps goes to the maintainers, not here.

What this does and does not show

Shown: a cheap-breadth allocation, pointed at a large codebase it had not scanned deeply before, surfaced a plausible, well-cited, easily-checked finding for $3 — alongside a wide spread of unverifiable and refuted leads that cost nothing further once triaged out.

Not shown: that this allocation generalizes — one scan on one target is one data point, and the n-day benchmark is the only number on this site meant to support a claim about recall across many cases. It is also not yet shown that this finding survives hands-on reproduction; "confirmed" here means the engine could not break its own claim, not that a human has — that step is in progress.