Mythos just moved the goalposts on software security.

On April 7, Anthropic announced a model it decided was too dangerous to release.

Claude Mythos Preview — a frontier model Anthropic won’t make generally available — spent a few weeks finding thousands of previously unknown, critical vulnerabilities in every major operating system and every major web browser. Anthropic engineers “with no formal security training” could ask Mythos to find remote code execution vulnerabilities overnight and wake up to working exploits. In one case, it found a flaw in a line of code that had been tested five million times without anyone spotting it. In another, it surfaced a bug in a 27-year-old operating system widely regarded as one of the most secure ever built.

In response, Anthropic launched Project Glasswing — a restricted consortium of AWS, Apple, Google, Microsoft, CrowdStrike, Palo Alto Networks, Nvidia, and a short list of banks — to quietly patch the world’s foundational software before this capability leaks to attackers.

Anthropic’s co-founder Jack Clark has publicly stated the obvious: this is not a special one-off. Similar capabilities will show up in other labs’ models in months, and in open-weight Chinese models within 12 to 18 months.

If you build, sell, or depend on software, the clock just started.

Why governments and banks are panicking

Bank of England Governor Andrew Bailey warned at Columbia last week that Mythos could “crack the whole cyber risk world open.” The US Treasury Secretary and Fed Chair convened an emergency meeting with the CEOs of JPMorgan, Goldman Sachs, Citigroup, Bank of America, and Morgan Stanley to talk through what to do. South Korea’s Financial Supervisory Service and European banking supervisors are running parallel conversations.

The anxiety is specific and well-founded:

• Banks run decades-old code under modern wrappers. A typical large bank’s stack blends mainframe COBOL, mid-era Java, cloud-native microservices, and bolted-on SaaS. Every seam is a candidate for exactly the kind of bug Mythos is good at finding.
• The financial system is consolidated onto a handful of cloud providers. A single Mythos-class exploit against AWS, Azure, or GCP cascades across the whole regulated banking system simultaneously.
• Government agencies inherit the same attack surface — the same operating systems, the same browsers, the same upstream open source dependencies — without the budgets or talent pools the banks can throw at the problem.

So the tier-one banks are getting early Glasswing access. The big tech companies are patching. Regulators are running scenarios. The systemic-risk story is, for now, being actively managed.

Why everyone else should be paying closer attention, not less

Here’s the part that’s getting lost in the “will it crash the banking system?” coverage.

Mythos-class capability doesn’t just find new bugs. It invalidates the implicit security posture of every piece of software that was previously “tested fine.”

The UK’s AI Security Institute was explicit about this: evaluation frameworks that used to meaningfully differentiate between safe and unsafe systems no longer do, because the cost of thorough review has collapsed. A code path that survived five million test runs isn’t “battle-tested” anymore — it’s just unexamined by the right reviewer. Every static analysis pass, every penetration test, every SOC 2 audit, every “we had a security firm look at it in 2023” is now working from a bar that attackers can clear.

The Veracode 2025 benchmark found that the average organization carries 100,000+ open vulnerabilities with a 252-day mean time to remediation. That math worked when attackers needed weeks to weaponize a new disclosure. It does not work when the weaponization timeline drops to hours.

And this matters most for the businesses who are least ready for it: specialized software companies — health tech, legal tech, government tech, fintech vendors — that handle sensitive data without a 200-person security team.

If you’re a health tech company (or any specialized software vendor without a big security org), this is your moment

I work with a healthcare-adjacent company (SiteLabs, activating independent pharmacies for clinical trial recruitment) and I talk to a lot of founders in similar positions: deeply regulated data, real consequences for a breach, and a security team that is either one part-time contractor, the CTO’s weekend, or both.

A few things are true at once right now:

1. The proposed 2025 HIPAA Security Rule amendments will mandate encryption, MFA, and network segmentation for all covered entities and business associates — expected to be finalized this year. Enforcement under OCR’s Risk Analysis Initiative is already up, with 11 actions by early 2026.
2. The Change Healthcare breach — 190+ million patients, $2.4B+ in costs to UnitedHealth, catalyst for the new HIPAA rules — was caused by a single Citrix portal without MFA. A Mythos-class attacker would have found that in minutes.
3. The average healthcare breach now costs $7.42M, and healthcare has been the costliest industry for 14 consecutive years. A breach-caused downtime runs roughly $900K per day.
4. You will not be in Project Glasswing. Neither will your EHR vendor, your patient intake tool, or your clinical trial recruitment platform. The defensive version of this capability is rolling out top-down to the biggest names first.

So what do you actually do?

Stop treating security as a compliance checkbox and start treating it as an architectural commitment. The good news is that the highest-leverage moves are unglamorous and affordable:

• MFA everywhere, non-negotiable. Including on admin consoles, cloud infrastructure, and every SaaS tool that touches PHI. This one control would have prevented Change Healthcare, the single largest healthcare breach in U.S. history.
• Ruthless patching discipline with a defined SLA. Pick a number — 15 days for critical CVEs is a reasonable benchmark — and actually hit it. This is more about operational consistency than tooling. Automate dependency updates (Dependabot, Renovate, Snyk) and commit to reviewing them weekly.
• Know your software bill of materials. You cannot patch what you cannot see. Every open source dependency in your stack is now in Mythos’s scope. An SBOM and a process for tracking CVEs against it is a weekend project that pays for itself the first time a Log4j-style event happens.
• Network segmentation and least privilege, done in code. If your app server is compromised, it should not have credentials to dump your entire patient database. Identity and access controls are where real resilience lives when prevention fails.
• Tested backups with a 72-hour restore target. Not “we have backups.” Actual restore drills. Ransomware operators using Mythos-class capability will be faster and more sophisticated than what most disaster recovery plans were designed for.
• Buy a vCISO, not a full security team. Fractional CISO services and managed detection and response (MDR) providers are the single highest-ROI security spend for a 5–50 person software company right now. You are not going to out-hire Goldman Sachs for security talent. You can rent adult supervision for the cost of a senior engineer.
• Use AI defensively, too. The same models that can find vulnerabilities at offensive speed can find them for you, at defensive speed. Automated code review, secret scanning, and vulnerability triage are all radically cheaper and better than they were 18 months ago. Build this into your CI/CD pipeline now.
• Get your vendor risk house in order. Your BAAs, your subprocessors, your third-party integrations — every one of them is a potential entry point, and regulators will hold you accountable for their hygiene.

None of this is new advice. What’s new is that the window to implement it is closing fast, and the downside of not doing it has gone up by an order of magnitude.

The uncomfortable reframing

Mythos is a warning shot. The loud version of the story is about whether AI-assisted attackers can take down the banking system. The quiet version — the one that matters for most of the software industry — is that the baseline for “secure enough” just moved, and the organizations that refuse to move with it are going to find out the hard way.

If you run a specialized software company, you don’t need to build a Fortune 500 security program. You need to get the fundamentals actually right, automate what you can, rent what you can’t, and stop assuming that “it hasn’t been a problem yet” is a forward-looking statement.

The next year is going to separate the teams who treat this seriously from the ones who don’t. You don’t want to find out which one you are by reading about yourself in a breach disclosure.

---

If you’re running a health tech, fintech, or other regulated-data software company and want help thinking through your AI and security posture, that’s the kind of thing I do at Stealthy Good AI. Let’s talk.