Anthropic's GitHub scanner goes Enterprise-only, and the model behind it is not established

Anthropic's description of the scanner is about scope rather than cleverness: it traces data flows, reasons about how components interact, and looks for vulnerabilities whose exploitability depends on the surrounding application rather than an isolated line of code ^[10]. Candidates pass through multiple stages of self-verification before they appear in the product ^[11], and each surviving finding arrives with location, impact, reproduction steps, severity, category and a recommended remediation ^[12]. Then it stops being the tool's problem. The output is a recommendation for human review, and the proposed fix opens in Claude Code against a model the customer already has in their account, according to Anthropic's product documentation ^[13].

That handoff is where the headline number should be divided. Anthropic says Mythos Preview and roughly 50 early partners identified more than 10,000 high- or critical-severity vulnerabilities ^[17]. That is upwards of 200 per partner ^[19], and RuntimeWire's caveat is the part that carries weight: this is model-reported findings at scale, not 10,000 independently confirmed and patched bugs ^[18]. Two hundred candidate criticals per organisation is a triage queue, and the queue is staffed by the buyer.

The naming deserves a slow read before procurement. The hosted Enterprise service and the Claude Security plugin share a name but not a deployment model, with Enterprise scans running through Anthropic's platform while the plugin runs locally on the Claude access the customer already pays for ^[9]. Separately, an official Claude account post describes Mythos as the model scanning repositories, and its publication date is not established in the supplied capture ^[4]. The sources do not settle whether the hosted service later switched models, or whether that post describes a different deployment entirely ^[5].

Chronology does not rescue the reader either. The Enterprise public beta opened in late April 2026 ^[1], roughly six weeks before Anthropic announced Mythos 5 on June 9 as an upgrade to Mythos Preview for approved Project Glasswing partners ^[6][21]. A post that says "Mythos" therefore does not tell you which Mythos, or which product surface it was scanning from. The one thing the material does establish is the direction of gating: Anthropic calls Mythos 5 its most capable model for cybersecurity and biology research and restricts it to a small group of vetted partners on the grounds that the same capabilities could be used to build cyberattacks or dangerous weapons ^[7].

The pricing sits in a different world too. Mythos trusted-access rates start at $10 per million input tokens and $50 per million output tokens ^[8], a five-to-one output premium ^[20] that reads like a model whose expensive work is writing, not reading. Enterprise scans, by contrast, bill as direct token costs with no separate Claude Security platform fee ^[14]. RuntimeWire's reading is that discovery generates model usage while remediation sends the engineer into Claude Code, leaving Anthropic in possession of both ends of the workflow ^[15]. The honest procurement assumption is the dull one: you are buying a hosted workflow that runs some Claude model, metered as consumption, with the verification labour on your side of the line.