Tufin Global Field CTO Erez Tadmor sat down with Jeffrey Spear, Tufin’s CISO, for a no-fluff hype check on Anthropic’s Mythos, Project Glasswing, and the Fable reversal.
There’s been a lot of noise lately around Anthropic’s Mythos and Project Glasswing — and if you’ve opened LinkedIn in the past few weeks, you already know what we mean. Every vendor has a hot take. Every headline feels urgent. And then, in the space of a single week, the story got sharper: Anthropic released Fable 5 — the public, “safer” version of its most powerful model — and the US government ordered it pulled three days later. So we sat down for a frank, no-fluff conversation about what it actually means for security teams. Or, as Tufin CISO Jeffrey Spear put it: “a little bit of mythos versus reality.”
So, What Is Mythos — and Why Does Glasswing Matter?
Mythos is Anthropic’s latest frontier model. “All of the vendors in the space have some type of frontier model,” Spear notes — but what’s different here is that Anthropic launched Project Glasswing alongside it, an initiative that gives selected industry partners early preview access, specifically so defenders can find and fix vulnerabilities before these capabilities are broadly available.
The idea is sound: get defenders ahead of the curve. But the question everyone’s really asking is whether this represents a genuine step-change in how we approach vulnerability management, or whether it’s mostly marketing noise. As Spear sees it, that answer is “getting uncovered day by day at this point.”
“It’s honestly not changing the game, but it’s collapsing a lot of the time and effort that went into parts of the game.” — Jeffrey Spear, CISO, Tufin
Here’s the honest take: Mythos isn’t reinventing security. What it’s doing is collapsing the time and effort that vulnerability research used to require.
Previously, a skilled researcher needed extended access to a codebase, dedicated tooling, and often a supporting team to test potential issues. That whole process can now be handed to a model. You point it at a codebase, ask it to find novel vulnerabilities and unexpected interactions, and it returns a list — fast. That speed is exactly what’s unnerving security leaders.
And that anxiety isn’t hypothetical — it’s what Spear is hearing directly from the field. “That’s where a lot of the concern, at least in our client base, is coming from,” he says. “What happens when you get 10 to 20 critical vulnerabilities? How are you going to address those with the efficiency and speed that’s required?” It’s not that the type of threat has changed. It’s that most teams aren’t currently built for that volume at that pace.
The Time-to-Exploit Window Keeps Collapsing
We’ve watched the window between vulnerability discovery and active exploitation shrink from months to weeks to days. With AI-assisted attack development, that compression is accelerating. “There was a term for it — zero-days — where vulnerabilities nobody knew about were already being used to attack people,” Spear says. “It’s even less than zero days now.”
That compression changes the calculus for every CISO. Patching fast used to be a competitive advantage. Now, as Spear puts it, “there are going to be instances in the industry where vulnerabilities come out, and there will not be patches, because we can’t make them fast enough.”
When the “Safe” Version Lasts Three Days
If you want a clean illustration of how thin certainty is right now, look at what happened to Fable.
Anthropic had already declined to release Mythos to the public, on the grounds that it was too capable at hacking — which is why it went to a small set of partners through Glasswing rather than to everyone. Fable 5 was supposed to be the answer to that problem: the same underlying model, wrapped in added safeguards designed to stop it being used for cyberattacks. The guardrails classify requests as safe or unsafe and route the risky ones to a less capable model. It launched publicly on June 9.
It was gone by June 12.
The US government issued an export-control directive after concluding that someone had found a way to jailbreak Fable — to bypass those safeguards and pull out information useful for cyberattacks. Because the order reached foreign nationals everywhere, including Anthropic’s own staff, the only way to comply was to switch the models off for every customer worldwide. Anthropic complied the same day, while publicly disagreeing with the decision — arguing that a narrow jailbreak shouldn’t force the recall of a model already in wide use, and that applying that standard across the board would halt frontier releases industry-wide.
You can take a view on who’s right in that dispute. But step back from it and the lesson for defenders is hard to miss: the model built specifically to be the safe, deployable one was shown to have a bypass surface within days of release, and it was offline within 72 hours of going live. Guardrails on these systems aren’t bulletproof; the people probing them are fast, and availability itself can evaporate overnight for reasons that have nothing to do with your environment.
Which leads to a simple principle. When the situation is this uncertain — when the “safe” option is here one day and gone the next — you fall back on the controls you actually own and can act on yourself.
The Network Is the Control You Already Own
The practical response to Mythos isn’t to rush out and find the shiniest new AI security tool. It’s to get serious about the fundamentals — and the network is at the center of that.
“The network is the most low-level defense that you can implement almost universally. If you can control what can access it and what it can speak to, that is a very effective defense for even a novel vulnerability.” — Jeffrey Spear, CISO, Tufin
Think about what the network can actually do for you in this environment:
Reachability and exploitability are different problems. You probably already have a long list of vulnerabilities. The question isn’t whether AI will find more of them — it will. The question is: which of those vulnerabilities are actually reachable? Can an attacker get to them through your firewalls, security groups, and network controls? That’s a network question, and it’s one that can dramatically shrink your real attack surface.
The network can contain what you can’t yet patch. When a critical vulnerability drops and no patch exists — or your patch cycle is weeks out — network segmentation and access controls can block the path an attacker would need to exploit it. The network is a defense layer you can act on immediately, regardless of vendor timelines — or, as the Fable episode showed, regardless of whether a given model or tool is even online tomorrow.
Visibility into risky access is non-negotiable. Overly permissive rules, broad access that was never cleaned up, traffic paths that shouldn’t exist — these are exactly the kinds of gaps that AI-assisted attackers can now find and exploit much faster than before. Knowing where those gaps are, before an attacker does, is foundational work.
This is the work Tufin has specialized in for years: analyzing reachability across firewalls and cloud security groups, surfacing overly permissive access, and validating segmentation — so the network can serve as the control layer when patching can’t keep up.
The CISO Calculus Right Now
“It used to be securing against attackers,” Spear says. “Now it’s securing with AI and securing against AI. Every CISO is doing that calculus.”
The answer to both sides of that calculus starts with the same thing: understanding your network. Knowing what’s exposed, what’s reachable, how traffic actually moves through your environment, and where you can use the network itself as a control when other options aren’t available.
Mythos is a signal, not a surprise — and the Fable reversal is the exclamation point. Together they confirm a direction the industry has been moving toward for years: AI will increasingly be part of both the attack and the defense, the capabilities will arrive faster than the safeguards around them settle, and access to those capabilities can shift without warning. The organizations that fare best will be the ones that invested in that foundational visibility now, rather than scrambling to react when the next wave arrives.
Three Things to Do This Week
- Verify reachability of your known criticals. Before AI adds to the pile, confirm which of your existing critical vulnerabilities are actually reachable through the network — and prioritize those.
- Audit overly permissive access. Find the broad rules and stale access paths that were never cleaned up — they’re the gaps an AI-assisted attacker will find first.
- Validate segmentation as a containment plan. Confirm you can block an exploit path with network controls today, so you have a real option when the next unpatched vulnerability drops — or the next tool you were counting on goes dark.
Speed is the name of the game. The network is how you play it.
Ready to Learn More
Get a Demo
