ICE's Mythos deployment gives Project Glasswing a market-infrastructure test (2026)

ICE is already using Claude Mythos Preview inside financial-market infrastructure, not just evaluating it from the sidelines. A HousingWire interview published Jun. 19 says Intercontinental Exchange has been using Anthropic's Project Glasswing tooling for several weeks, with the deployment covering vulnerability discovery, validation, and remediation workflows. 1

This is a follow-up brief, not a new-join announcement. ICE officially announced on Jun. 3 that it had joined Project Glasswing and deployed Claude Mythos Preview across its exchanges, clearinghouses, data services business, and mortgage technology platform, including the New York Stock Exchange. 2 The new information is operational: ICE executives are now describing how the model is being used and how they plan to judge whether it is worth the risk.

The fresh event sits on a two-step timeline: ICE announced the deployment on Jun. 3, then described active use and measurement plans in a Jun. 19 interview. 1 2

ICE x Project Glasswing at a glance

Public facts from the ICE announcement and follow-up interview

Official deployment announcement	Jun. 3
Fresh operating interview	Jun. 19
Publicly named scope	All ICE businesses

Loading stats card…

Why this customer matters

ICE is a stronger proof point for Anthropic than a generic enterprise logo. The company runs exchanges and clearing infrastructure, owns the NYSE, and sells mortgage-technology systems that touch regulated financial workflows. 2 That makes the deployment a test of whether Anthropic can place a restricted, security-focused model inside systems where auditability and operational controls matter as much as raw model capability.

Signal	What is confirmed	Why it matters for Anthropic
Deployment scope	ICE says Claude Mythos Preview is deployed across all ICE businesses, including the NYSE, exchanges, clearinghouses, data services, and mortgage technology. 2	Project Glasswing is reaching critical financial infrastructure, not only software companies or lab pilots.
Governance owner	ICE says it manages the deployment, security architecture, and governance of the model's use. 2	The customer, not only Anthropic, is taking responsibility for control design inside regulated operations.
Fresh operating detail	ICE CISO Steve Pugh told HousingWire that ICE was invited into one of Project Glasswing's waves and had been using Mythos for the last few weeks. 1	The program has moved from announcement language into active customer use.

How ICE says it is using Mythos

Pugh framed the model as a speed and triage tool. He said AI models are not necessarily creating new exploit categories, but they compress the time needed to find and act on weaknesses. Mythos, in his description, can scan broadly and quickly decide whether something is merely a defect or a vulnerability that could lead to exploitation or system control. 1

ICE is not saying it will automatically fix every model-reported issue. Pugh told HousingWire that Mythos can create a proof-of-concept exploit for a vulnerability, then run an adversarial check against findings to reduce false positives. He said the result is meant to be a tighter view of which vulnerabilities are real and which may be false positives. 1

What ICE says it will measure

The three evaluation questions Pugh described in the HousingWire interview

Finding quality	False positives
Detection speed	Faster than tools
Customer benefit	Time to fix

Loading stats card…

That matters because Anthropic's own Mythos Preview research argued that the model can identify and exploit serious vulnerabilities, including zero-days, and that Project Glasswing was created to put those capabilities in defenders' hands before similar capabilities become broadly available. 3 ICE is now one of the highest-profile tests of that defensive thesis.

What ICE will measure

The most useful part of the HousingWire interview is the metric stack. Pugh said ICE is looking at whether Mythos finds the findings the company wants to know about, whether it finds them faster than existing tools, and whether ICE can turn those findings into fixes quickly enough to benefit customers. 1

That gives Anthropic three customer-facing proof points to watch:

False-positive control: whether model-generated findings survive human and adversarial validation.
Detection speed: whether Mythos beats ICE's existing static-analysis, application-security, and penetration-testing paths on material issues.
Remediation throughput: whether the model shortens the time from finding to fix, not just the time from prompt to report.

Bob Hart, president of ICE Mortgage Technology, also said some customers contacted ICE after the Mythos press release to thank the company for joining the program, while security questions from customers were still emerging more slowly than AI-governance questions. 1 That suggests the commercial value may depend on whether buyers start asking for concrete AI-security posture, not just AI-use policies.

What remains undisclosed

ICE and Anthropic have not disclosed contract value, duration, number of systems covered, model access terms, or whether the deployment changed after the later U.S. export-control actions around Fable 5 and Mythos 5. The public material also does not quantify how many vulnerabilities Mythos has found inside ICE systems, how many were true positives, or how many have been remediated. 2

The near-term watch item is whether other regulated Project Glasswing participants start publishing the same kind of operating metrics. One named financial-infrastructure customer is useful for Anthropic. Repeatable evidence across several customers would be stronger: fewer false positives, faster validated discovery, and measurable fix rates inside production security programs.