The most safety-focused AI company just had two operational governance failures.

The most safety-focused AI company in the world just had two operational governance failures in the same week.

Not exploits. Not hacks. Misconfigurations.

The first: draft materials about an unreleased AI model (the one now public as “Claude Mythos Preview”) sat in a publicly accessible data store. Anyone who knew the URL could read them.

The second, a few days later: a build configuration shipped a source map for Claude.ai, the consumer product. A source map turns compressed code back into something readable. The exposed map revealed the internal module structure of a flagship AI product, including names of subsystems most users will never know exist.

Different systems. Different mechanisms. Same company. Same week.

Then this is where it gets interesting. Cybersecurity leaders running Mythos against their own code have found a second gap. The model is genuinely good at surfacing real vulnerabilities. But, the patches it proposes are weak and still need humans to check, refine, and ship. AI works on the input side. Process still has to do most of the work on the output side.

What’s the takeaway for you?

Anthropic publishes a constant stream of AI safety research. Their CISO is publicly sprinting on nation-state-level internal defenses. The discipline is real.

And the basics still got missed. Twice.

This isn’t an Advanced Persistent Threat. It’s a Basic Persistent Failure. At one of the most disciplined AI companies in the world. Back-to-back.

Here’s the takeaway:

You’re not building AI. You’re buying it. Copilot. ChatGPT Enterprise. AI features bundled into SaaS tools you already use. You don’t control the plumbing underneath. You trust the vendor to get it right.

If Anthropic, with all of its resources and culture, ships a build pipeline that publishes a source map by accident, what’s running inside the smaller AI vendors your team is integrating with this quarter?

The AI works. The processes around it don’t keep up.

That’s the pattern. The risk of AI isn’t usually the AI itself. It’s the gap between how fast AI lets a company move and how fast its operational discipline can match. Anthropic’s CEO has said the majority of the company’s code is now written by Claude. Engineers ship multiple releases per day. The velocity is extraordinary. The mundane stuff, like which files get included in a software package, didn’t keep up.

Your organization has the same gap. Not in how you build software, but in how fast you’re adopting AI tools. Employees connect AI assistants to production data. Teams adopt AI-powered SaaS without IT review. Departments build workflows on AI tools that nobody in security or compliance knows about.

Here’s what you can do this week.

1. Ask your AI vendors what happens when their system crashes mid-task with your data in memory. If they can’t answer clearly, that’s a finding.
2. Audit what AI tools are actually connected to your environment. Not the ones you approved. The ones that are there. Check OAuth integrations, API keys, browser extensions, and payment records from Finance.
3. Review the default permissions on every AI tool you’ve deployed. Most ship wide open to reduce onboarding friction. If nobody tightened them, you’re running with unlocked doors.
4. Update your board conversation. The old question was “are we secure?” The better question is “is our AI adoption speed outrunning our ability to govern what we’re adopting?” This is also the moment to ask for what you need. Budget. Headcount. A governance miss at the most safety-conscious AI company in the world is the argument you’ve been waiting for.

The lesson isn’t about Anthropic. It’s about the pattern. Every organization adopting AI is moving faster than its governance can follow.

One more thing the news cycle is getting wrong. The story is that only Mythos can do this work. That’s not what independent researchers are finding. A team at AISLE reproduced Mythos’s flagship vulnerability discoveries using small, open-weights models, including one that costs about eleven cents per million tokens. The frontier capability is already widely available. The moat is the system around the model, not the model itself.

There’s a bigger version of this pattern just ahead. Anthropic’s CISO recently told a private group of cybersecurity leaders that “two years from now, we will be in a world where if you ship a security vulnerability, it will be exploited immediately.” Not in days. Immediately. Your patch management program was built for a world with weeks between discovery and attack. That world is changing fast.

The AI worked. The question is whether your process will.

How fast is your organization adopting AI compared to how fast you’re governing it?

-Kip