Let's talk security strategy
Cybersecurity and compliance are critical to the success of any business. Protect sensitive information, lock down your assets and give your team the tools and training they need to outsmart threat actors.
Contact Us
Milo is a crypto-backed mortgage lender. Its platform has written over $250 million in total loans and more than $100 million in crypto mortgages, across more than 5,000 applications. A lender that holds borrowers' financial data and digital assets has to prove its security, not just describe it. For four consecutive years, Milo has done exactly that: SOC 2 with no exceptions. IronCloud has been its security and infrastructure partner throughout, and the program has grown stronger with every year of the relationship.
Engineering at Milo is led by Ricardo Morales, SVP of Software Development, who oversees platform development and the scaling of Milo's cloud-native architecture. His team runs AI in production across the business, from document classification at intake to an AI-powered sales development outreach system and an analytics assistant on the data warehouse. It is a fast-moving engineering organization, and its security program was built to keep pace with it.
Milo first engaged IronCloud for infrastructure services. As the platform grew, the two teams made a deliberate choice: security would be built into that foundation, not layered on afterward.
IronCloud built Milo's AWS environment from the ground up, wrote its security policies, built the CI workflows that carry code to production, and brought the platform under IronCloud's monitoring services for logging, monitoring and alerting. Because the same team designed the systems and the controls at the same time, the controls described how Milo actually operated from the first day. Three months in, Milo was SOC 2-ready.
SOC 2 Type II is not a one-time exam. Every year, auditors test whether controls operated effectively across a full review period, measured against the Security trust services criteria. Four years in, Milo has no exceptions on record.
That consistency comes from continuity. IronCloud still runs the infrastructure, CI workflows, monitoring and policies the auditors examine, so the program grows with the platform instead of being rebuilt before each audit. Each year IronCloud also supports Milo's evidence collection, including proof that penetration test findings were remediated.
IronCloud attacks what it builds before anyone else can. The testing is done by IronCloud's own team, never subcontracted, and it runs on two tracks. The first is steady pressure: recurring external penetration tests and web scanning against Milo's public-facing infrastructure, with every finding tracked until it is fixed.
The second is the annual red team engagement, and nothing about it comes from a playbook. Each one is a multi-stage operation designed from scratch around the threats Milo faces that year. IronCloud writes the scenarios, pretexts and attack paths itself, never reuses them, and never repeats them for another client. No commercial tool or platform produces this kind of work, because the attack is built for one organization at one moment in time.
This year's engagement went straight to the top. IronCloud set out to answer a question no firewall can: would Milo's leaders act on a call that sounded exactly like their own CEO?
An off-the-shelf voice clone would not survive that test, because the people on the other end hear the real voice every day. So IronCloud trained its own models on the CEO's voice and kept refining them until the clone carried the CEO's pitch, tone, pace and intonation, the small patterns colleagues recognize without ever thinking about them.
Then IronCloud took people out of the loop entirely. Every call was run by AI agents IronCloud trained for this engagement, speaking in the CEO's voice. When someone picked up, the agent held a live conversation. When no one did, it left a voicemail that sounded like the CEO calling back between meetings. The agents were trained for the unscripted parts of a real call, too: the pause before an answer, the interruption, the moment of doubt. They were built to work a call the way a skilled social engineer would: read the response, handle objections, stay assertive, and turn up the pressure when someone hesitated. The calls were coordinated across the leadership team so the story held together from one conversation to the next. No IronCloud operator was on the line. From the first ring to the last word, it was a 100 percent AI-driven attack, and every leader who answered heard their CEO.
The voice asked for exactly what a real attacker would want: approve payments and reset access. One leader did what a strong control looks like in practice. They ended the call, dialed back on a known number, and stopped it.
The engagement did not end with the calls. IronCloud ran a tabletop exercise with Milo's leadership and followed it with training on how to recognize and handle vishing. Milo's leaders trained against an attack that had just been run on them, not a hypothetical one, and the lesson was concrete: when a voice can be generated, a familiar voice is no longer proof of who is calling. Only verification is.
That is why IronCloud builds every red team from scratch. Attackers do not reuse last year's playbook, and neither does IronCloud. This year, the threat that mattered was an executive's voice turned into an attack tool, so that is what IronCloud built. Every engagement is sanctioned and scoped with the client, and this one was planned together with Ricardo Morales and Milo's leadership team. Anything beyond what is shared here stays between IronCloud and Milo.
Milo's most recent penetration test used Nemesis, a security scanning platform IronCloud built and runs itself. It is unrelated to the open source project of the same name. Nemesis maps a client's external attack surface and cloud configuration, covering DNS, exposed services, TLS, web applications and AWS accounts, and checks the software it finds against known CVEs.
AI agents do the thinking inside Nemesis, and guardrails decide what they are allowed to do. An orchestrating agent plans each scan and changes course as results come in. A reasoning agent reads discovery output as it streams in, decides what deserves a closer look, and proposes follow-up investigations, the way an experienced tester would. Before any of those tasks runs, a safety gate checks it: only known, non-destructive actions are allowed, run time is capped, anything that could change a target is blocked, and nothing touches a system the client has not authorized. Cloud access is read-only by design. Critical and high-severity findings then go to a verification agent that records its confidence and the evidence behind each conclusion, so the people reviewing a finding can see exactly why the system believes it. The result is offensive security work that moves at machine speed and stays inside its mandate.
Findings are ranked by how exploitable they are in practice, not by severity alone, using CVSS, EPSS exploit probability, the CISA Known Exploited Vulnerabilities catalog and public exploit availability. Each finding keeps one history across scans: when a fix ships, a follow-up scan shows it resolved, and anything that comes back shows as regressed. That history is what turns a penetration test into audit evidence, a record that ties each finding to its fix.
“They have transformed our business with secure, efficient and cost effective cloud solutions tailored to our needs.”
Results approved for publication by Milo.
Published November 6, 2026
Cybersecurity and compliance are critical to the success of any business. Protect sensitive information, lock down your assets and give your team the tools and training they need to outsmart threat actors.
Contact Us