Anthropic’s new initiative is giving cyber defense a durable advantage

POV by Mike Hulbert, VP of Sales

I remember reading an AI fictional scenario only a few years ago where an AI fake video maker was pitted against an adversary AI until the video they made could outsmart the AI attacker’s ability to detect it as fake.

Fast forward to April 2026, and you may have noticed the current top AI concern is several orders of magnitude higher:

AI now has the capability of independently finding and exploiting highly critical vulnerabilities in software and infrastructure.

The model in question so far is Anthropic’s new Claude Mythos Preview. The company announced in early April that it would not yet release this model because they had serious concerns about its offensive cyber capability.

To put this in context and to see where I’m taking these thoughts, the company’s previous flagship model was good at finding vulnerabilities but far less good at exploiting them.

Mythos Preview Is leading the new order of AI ability. Not intentionally, I understand. But as a result of its improved abilities in coding, reasoning, and completing autonomous tasks.

However, for me, the real headline here is this: Anthropic voluntarily held back this model from public release.

That’s a top AI company putting safety over speed-to-market – and it sets a precedent the rest of the industry will have to respond to, whether they want to or not.

What’s Being Done to Safeguard Business Systems From These Super-Intelligent (and Potentially Business-Threatening) AI Models?

Anthropic has decided to put AI-driven cybersecurity first. They’ve invited a group of all the main cloud provider and network security companies to work with them defensively in what they’re calling Project Glasswing.

It’s a collaborative industry initiative to secure critical software infrastructure against the evolving risk of AI-led cyber threats.

For example, one of the group is Cisco – the largest supplier of network equipment in the world. They’re now integrating Mythos Preview into their security and development workflows.

My extensive experience in delivering safe IT services tells me this is a sound move.

Just like we now have tools to detect viruses, email spam, and phishing, the big players in the industry will need to develop new tools that ward against new security vulnerabilities.

The Project Glasswing partnership has been granted $100 million in “committed usage credits” to achieve a defense solution before a public release allows malefactors to misuse Mythros Preview’s ability.

I believe Anthropic’s pivot to “deploy it defensively first” is smart. Recent challenging developments in AI require this kind of strategic approach.

But it also confirms what a lot of security pros have suspected: The offensive capability of these models is serious enough to warrant this high level of caution.

It’s crucial that all security defenses evolve in parallel with the abilities of the new AI models. It’s the only way to maintain the integrity and safety of the internet for all of us.

However, there are issues for mid-market clients that will need to be addressed.

What This Means for Mid-Market Clients: The Need for Remediation Services

From my perspective, overseeing our service provision every day, the gap between enterprise and mid-market security is about to get wider.

Large players like Cisco get early access to AI-powered defense tools (for the good reasons I’ve mentioned); but smaller companies are still patching manually and reactively.

That asymmetry produces a real exposure risk.

And of course, vulnerability discovery is getting faster on both sides. If defenders are finding flaws faster with the help of the latest AI, so are attackers.

Holding back the release of one model (as Anthropic has done) won’t stop attackers using other similar models to discover and exploit vulnerabilities in business systems.

It would be naive to assume Mythos Preview is a one-off. The last few years have provided ample proof that models leapfrog each other in capabilities almost weekly.

Remediation will therefore be the area that matters most for mid-market clients.

Those who aren’t investing in faster remediation workflows – not just detection – could find themselves worse off even as the tools improve.

Where is This AI-Driven Cybersecurity Narrative Going in the Near Future?

Anthropic’s delayed-release narrative won’t stay a “big vendor” story for long. It’s not just about prioritizing AI-driven cybersecurity

I believe that in 18–24 months, AI-assisted vulnerability scanning will be table stakes – and the differentiator will be who can act on findings quickly. As I said, remediation will be key.

The bottleneck will quickly shift from finding problems to fixing them – which is mostly a people and process problem, not a technology one.

Another thing I foresee in the near future is this – and it’s significant…

The Anthropic/big partner model hints at a future where AI vendors are essentially co-signers of cybersecurity outcomes.

That’s a very different kind of vendor relationship than the industry is used to.

What the Future Vendor Relationship Will Look Like

In the future, companies that sell AI tools won’t just provide software and walk away. They’ll share responsibility for what happens with it, especially around AI-driven cybersecurity.

Right now, most vendors just supply a product, and the customer is largely responsible for using it safely.

With this concept of “co-signer of cybersecurity,” the AI vendor is more like a partner who’s partly accountable if something goes wrong (like a breach caused by their system).

This implies:

• Shared liability for cybersecurity outcomes
• More guarantees or assurances from AI vendors
• Deeper, ongoing involvement rather than a one-off sale

This will be – for some mid-size businesses – a very different relationship. It shifts vendors from being suppliers to being accountable partners.

So vendor partnerships will matter when you select a technology partner. You need to rely on Tier 1 solutions to ensure speed and safety.

Why Katalyst?

This is exactly the kind of accountability gap that mid-market organizations can’t afford to leave open, and it’s the gap Katalyst is built to close.

We’re a Digital Operations Partner for mid-market organizations, which means we don’t just advise on your security posture and hand you a report. We design it, operate it continuously, and take full accountability for what happens day to day. That includes the Protect pillar: identity controls, endpoint monitoring, tested incident response, and a security posture you can demonstrate to auditors and your board, not just a list of tools you’ve purchased.

With the advent of AI that can autonomously attack your systems, like Anthropic’s Mythos Preview, this is non-negotiable in today’s world.

We optimize client experiences using people, process and technology. But the people and process elements are “non-AI must haves” when you want a secure digital backbone you can rely on.

On the remediation point specifically: Fast remediation isn’t just a workflow feature; it’s a function of having a partner who already knows your environment. When Katalyst manages your Digital Backbone, we’re not starting from scratch when a threat surfaces. We have the baseline, the playbooks, and the operational context to act. That’s the difference between a vendor you call and a partner who owns the outcome.

If you’d like to discuss any of these issues, schedule a call and ask for me.