AI is forcing the security industry to "think past detection", a lesson email account takeover has been trying to teach for years.
If you work in security, you've probably spent the last few weeks either reading about Claude Mythos or rolling your eyes at yet another vendor posting about it.
There's a real debate happening about how much of the Mythos story is genuine capability and how much is well-timed marketing. Bruce Schneier called it "mostly marketing hype." Sam Altman called it "fear-based marketing." The AISLE research team reproduced many of the headline findings using small, cheap, open-weight models — including one running at $0.11 per million tokens. When the CEO of HuggingFace can point at open-source models replicating what was pitched as a restricted breakthrough, it's fair to be skeptical about the packaging.
But regardless of where you land on Mythos itself, it has started an important conversation. A conversation about what happens when you can no longer assume detection will catch threats in time. And that's a conversation the email security industry has needed to have for years.
The conversation worth having
The specific model matters less than the trajectory. AI-driven vulnerability discovery is getting cheaper, faster, and more accessible — that's true whether the frontier is Mythos or something you can download from HuggingFace. The window between finding a vulnerability and weaponizing it keeps shrinking.
The industry's response has been to rediscover "assume breach." Not as a maturity milestone for sophisticated programs, but as a baseline operating assumption. I think that's right. But applying assume-breach to different domains requires different thinking and different tools. In appsec, it looks like runtime protection and software supply chain controls. On the endpoint, companies like CrowdStrike have been building containment and response for years — that's the whole premise of EDR.
In email, it looks like what we've been building at Material. And the connection between this broader "think past detection" moment and email security is one I haven't seen anyone make yet.
The founding bet
In 2016, John Podesta's email account was compromised. One credential, one account — and the resulting breach arguably changed the outcome of a presidential election.
We started building Material Security the following year. The conviction was simple: if you can influence an election by compromising a single email account, attackers will keep finding ways in. The entry point is too valuable, the attack surface too human, the payoff too high.
So the question we organized the company around wasn't "how do we catch more malicious emails?" It was: what happens after the attacker is already inside the account?
The question we organized the company around wasn't "how do we catch more malicious emails?" It was: "what happens after the attacker is already inside?"
In 2017, that was a hard sell. Email security was a detection market. Every vendor competed on catch rate. Nobody was asking what happened after the attacker got in.
What that looks like as a product
When an attacker takes over an email account, the damage isn't the phishing message that got them in. It's everything that comes after. Years of sensitive messages in the archive. The ability to impersonate a trusted colleague internally. Enough context to craft convincing BEC messages to anyone in the organization. All of it is accessible the moment the credential is stolen. No dwell time. No detection window.
We built containment controls for exactly this problem. We redact sensitive content in the archive so a compromised account doesn't automatically mean a compromised organization. We enforce step-up authentication on high-risk messages. We give security teams real-time tools to limit blast radius instead of reconstructing a forensic timeline after the fact.
If you've worked with products like Thinkst Canary or CrowdStrike's endpoint containment, the philosophy is similar: assume the attacker gets in, then design your architecture to limit what they can do with that access. We applied the same thinking to email, which needed it more than most people realized given how much sensitive data lives in the average inbox.
What we got wrong along the way
We also learned something important. When we started, we were so focused on containment that we underweighted inbound detection. Attackers still use email as the front door — phishing, malicious attachments, BEC — and pretending that was a solved problem was more ideological than practical. So we built detection capabilities too. Layers matter.
But what I keep seeing across the industry is that the mistake isn't investing in inbound detection. It's stopping there. Catch rate becomes the only metric. The entire email security strategy optimizes for one moment in a much longer attack lifecycle, and everything after a successful intrusion gets treated as an incident response problem rather than a design problem.
The mistake isn't investing in detection. It's treating detection as the entire strategy.
Email was already a zero-dwell-time problem
This is where the Mythos conversation connects to email in a way I don't think anyone has pointed out yet.
The whole debate — the real capability, the marketing, the AISLE reproductions — centers on exploit timelines compressing. The window between finding a vulnerability and weaponizing it getting dangerously short. That matters for AppSec and vulnerability management.
But email account takeover was already there. There is no patch cycle for a stolen credential. The moment an attacker has your password, they have your inbox, your archive, your ability to send as you. There is nothing to remediate. There is no window. The blast radius is immediate.
The AppSec community is arriving at "think past detection" right now because the Mythos conversation forced it. Email security should have been there already.
Containment isn't a phase anymore
For most of the time I've been in this industry, containment has sat at the far end of the security maturity curve. Every organization acknowledged it mattered. Most treated it as something they'd get to after the detection stack was optimized.
The broader trend — AI making offensive capabilities cheaper and more accessible — is making that sequencing harder to justify. The assumption that you'll have time to detect and respond before damage is done is getting less defensible every month, across every domain.
We started building for this in 2017. Not because we predicted any of this. Because one compromised email account told us everything we needed to know about what happens when containment isn't part of the architecture from the start.
