After Mythos: What Actually Changes for Cyber Risk

When Anthropic published its technical preview of Claude Mythos on April 7, the cybersecurity and insurance industries had one of those weeks where every conversation started with the same question. Is this the moment everything changes?

The short answer, from where we sit underwriting cyber risk at scale, is partly, and not in the way most of the commentary suggests. Mythos is a real inflection. It is not the end of cyber insurability. The distance between those two claims is where the actual work lives, for defenders, for security teams, and for the cyber insurance industry, which has been heading toward a confrontation with its own business model for a while now.

I want to walk through what Mythos actually changes, what it doesn't, why the popular framings (both "the sky is falling" and "we've been finding these bugs for years") miss the substance, and what any of it means for how cyber risk should be priced and managed going forward.

What's actually new in the Mythos claim

The technical heart of this matters, because much of the commentary has been sloppy about it.

Vulnerability discovery has never been the scarce resource in offensive cybersecurity. Fuzzers, static analysis, symbolic execution, human researchers, and earlier generations of LLM-assisted tooling find bugs in volume, every day. When cybersecurity veterans responded to Mythos by saying some version of "we have a pile of unfixed vulnerabilities already," they're correct. It's also the least interesting observation available on this topic.

The scarce resource has been weaponization, meaning the work of taking a crash or a memory safety violation and turning it into a reliable, repeatable capability against a real, patched system. That is a fundamentally different skill, and it has to get past a compounding stack of mitigations the industry spent two decades building.

ASLR (Address Space Layout Randomization) randomizes where code and data live in memory each time a program runs, so a bug that gives you execution control is useless without an independent information leak to tell you where to redirect it.

Control Flow Integrity, implemented via Intel CET, ARM Pointer Authentication, and clang's CFI, enforces that indirect calls land on valid function entry points. That kills most classes of code-reuse attack.

Sandboxing means that even full compromise of a browser renderer, a mobile app, or a document viewer lands you in a cage that can't touch the kernel or other processes directly. You need a second bug, usually a kernel or IPC vulnerability, to escape. Modern browser exploits are routinely three- and four-bug chains for this reason.

Modern allocator hardening, in the form of PartitionAlloc, hardened_malloc, and slab protections, makes the heap actively hostile to the grooming patterns that classic use-after-free and heap overflow exploits depend on.

Vulnerability discovery has never been the scarce resource in offensive cybersecurity. The scarce resource has been weaponization, meaning the work of taking a crash or a memory safety violation and turning it into a reliable, repeatable capability against a real, patched system.

Taken together, these mitigations are why a university student with a fuzzer can find a memory corruption bug in an afternoon, but turning that same bug into a reliable, one-click, cross-version remote code execution against a modern iPhone or Chrome-on-Windows has historically been a multi-week to multi-month project for a handful of specialized researchers worldwide. Those people earn six and seven figures doing it, at Project Zero, NSO Group, and the offensive divisions of nation-state programs.

What the Mythos technical report claims, and this is the part that matters, is that the model produced end-to-end capability, not just bug finding. Multi-vulnerability privilege escalation chains through the Linux kernel. A JIT heap spray escaping a browser sandbox. An autonomously authored RCE against FreeBSD. The UK AI Security Institute's independent evaluation put Mythos at 73% success on expert-level capture-the-flag challenges and reported it as the first model to complete a 32-step simulated corporate network attack.

If those results generalize, and "if" is doing real work there, what compresses isn't discovery. It's the weeks-to-months of specialist human engineering between "found a crash" and "own the machine." That is a materially different threat model than "AI finds CVEs," which has been true for years and didn't change anyone's risk posture.

What the skeptical case gets right, and where it falls short

When cybersecurity veterans push back on alarmist framing, they're working with good data. Most CVEs never get weaponized. Most weaponized ones don't yield remote, unauthenticated, pre-auth RCE. Most that do get patched before mass exploitation. Third-party testing of frontier models (AISLE's evaluation is worth reading in full here) has shown that detection-level capability is widely distributed. Even a small, inexpensive model can identify a critical FreeBSD issue correctly. Detection isn't the moat it's sometimes marketed as.

Where the skeptical case slips is in treating the discovery-versus-weaponization distinction as if both sides were equal commodities. They aren't. The commodity side, discovery, has been commoditized for a long time. The scarce side, engineering-grade exploit construction under modern mitigations, is what's now under pressure from Mythos-class capability. Collapsing those together and saying "we've always had this" is true for the wrong side of the equation.

The honest read, after spending time with Anthropic's published preview, the UK AISI evaluation, Fitch's commentary, and the claims data we see every day at Coalition, is that the ceiling of offensive capability moves less than the headlines suggest. The floor moves a lot. The set of actors who can produce a working exploit chain expands meaningfully. The time from disclosure to mass weaponization compresses. The economic constraint that kept most attacks targeted, namely skilled human labor being expensive, weakens.

Why the human bottleneck answer is only half right

A fair follow-up question. Isn't the real constraint on the offensive side still humans, specifically criminals with malicious intent, operational sophistication, and access to monetization infrastructure? Isn't that what actually bottlenecks global cybercrime?

Yes and no, and the distinction matters.

The historical bottleneck was a co-occurrence problem. You needed elite technical skill, patience for long exploit engineering, tolerance for legal and geopolitical risk, operational security to avoid attribution, a path to monetization, and strategic judgment about what to target. Those traits rarely clustered in the same person or team. Globally, the population that could actually convert a crash into a used capability was small. A few hundred freelance exploit developers. A few thousand inside nation-state programs. A handful of elite ransomware crews.

Mythos-class capability dissolves the technical skill constraint. It does not dissolve the others. Intent, monetization, legal risk, operational security, and strategic judgment remain human choke points. Crypto-laundering is still constrained by KYC regulation and chain analytics. Attribution and retaliation still happen at the state level. Most people still don't want to go to prison.

So the expansion isn't creating new criminals from nothing. It's taking the much larger population of people who already have the latent intent to cause harm and handing them capabilities previously reserved for the top of the pyramid. The "script kiddie" who could copy exploits but couldn't adapt them when something broke becomes meaningfully dangerous. Second-tier nation-states that historically couldn't field a competitive offensive program can. Insider threats get worse in interpretable ways.

Mythos-class capability dissolves the technical skill constraint. It does not dissolve the others. Intent, monetization, legal risk, operational security, and strategic judgment remain human choke points.

The economic scaling argument is worth being precise about, because it's easy to overstate. Criminals don't become indiscriminate simply because exploit construction gets cheaper. Mass, untargeted exploitation is noisy, attracts law enforcement attention, accelerates defender response, and often burns capability faster than it generates revenue. Sophisticated attackers have always been disciplined about operational security, and that doesn't change because a model can write an exploit for them.

What changes is the target selection math. Today, many potential targets are effectively off the table because the per-target cost of skilled exploitation exceeds the expected return. A mid-sized manufacturer with a moderate insurance policy. A regional hospital without a prestige attacker's attention. A municipal utility without a known ransom ceiling. These are economically uninteresting to a crew that has to spend human weeks per intrusion. If the per-target labor cost falls materially, the economically viable target set expands downward. You also get faster movement within the target sets a crew has already chosen, and a collapsing first-mover advantage for defenders who used to have time between CVE disclosure and active exploitation to patch.

None of that amounts to "scan and exploit everything." The tails of the distribution probably get heavier, the volume of mid-severity incidents probably rises, and the window defenders have historically relied on shrinks. That is a meaningful change in the shape of the threat landscape without being a change in its ceiling, and without contradicting the strategic logic attackers have always worked under.

What this means for cyber insurance, with appropriate caveats

The legacy cyber insurance product, characterized by an annual policy term, a point-in-time questionnaire, attestation-based controls, and static exclusion language designed to last a decade, was a workable bet in a world where the threat landscape evolved on a quarterly timeline. Mythos-class capability, and everything it implies about compressed weaponization windows, challenges the load-bearing assumptions of that product simultaneously. That much I am confident about.

Where the analysis gets harder is the systemic-event question. The common argument goes like this. Actuarial insurability rests on losses being roughly independent across insureds, so you can diversify across a portfolio and reinsure the tail. Cyber is weak on independence in principle, and CDK, Log4j, MOVEit, NotPetya, and SolarWinds are the canonical examples. Therefore, a world of AI-accelerated weaponization should make cyber increasingly uninsurable at the systemic tail.

The weakness in that argument is that it proves too much. In practice, the cyber insurance industry has absorbed those historical correlation events better than the abstract correlation argument predicts. Technology diversification is real. Not every system runs the vulnerable library. Not every vulnerable system is exposed to the public internet. Segmentation limits blast radius. Human attacker labor has continued to bottleneck mass exploitation even when the underlying vulnerability is ubiquitous. Defenders have moved quickly once a widely-exploited bug hits public awareness. Those events have been painful and loss-generating, but the line of business has continued and capacity has returned.

We don't know yet how much stress the absorption mechanism can take. The bet that historical correlation resilience continues becomes a more expensive bet to make on faith rather than evidence.

So the honest framing is not "correlation will destroy the cyber insurance line." It is that the industry's historical absorption of correlation events has relied on an implicit set of buffers. Slow attacker adaptation. Noisy exploitation attempts triggering early defender response. Human labor bottlenecking the speed of mass exploitation. Technology and configuration diversity limiting spread. What Mythos-class capability does, if the results generalize and if proliferation proceeds faster than defensive adoption, is stress-test those buffers. We don't know yet how much stress the absorption mechanism can take. The bet that historical correlation resilience continues becomes a more expensive bet to make on faith rather than evidence.

It follows that the answer to "what does the cyber insurance product look like in five years" is a genuine unknown rather than a foregone conclusion. Several outcomes are plausible.

One is that reinsurance capacity expands, cat bond structures mature, and the commercial market continues to write broad cyber coverage at somewhat higher prices with tighter control requirements. The market adapts through capital and discipline. Cyber cat bonds from Beazley, AXIS, and Hannover Re already show the structure can work.

A second is that the product bifurcates. A frequency layer, covering business email compromise, ransomware, privacy, and wire fraud, continues to look like today's cyber insurance. A systemic layer, covering widespread-vulnerability events, supply chain compromises, and nation-state-adjacent mass exploitation, moves behind explicit sublimits, or to government backstops, or to specialty carriers with different capital structures.

A third is that the commercial market proves more resilient than the skeptics assume, largely because defensive AI deployment moves faster than offensive AI proliferation, and the attacker-defender equilibrium stabilizes at a new level without a major product redesign.

Whichever path the product takes, the carriers best positioned to underwrite into it will be the ones with continuous visibility into their insureds' actual security posture.

Which path wins is not something I or anyone else can predict with confidence in April 2026. What I am confident about is this: Whichever path the product takes, the carriers best positioned to underwrite into it will be the ones with continuous visibility into their insureds' actual security posture. Static, attestation-based underwriting is running out of road regardless of which market outcome emerges.

Where traditional underwriting runs out of road

The specific place the legacy model breaks is underwriting itself. For the last decade, most cyber policies have been underwritten on an annual renewal cycle, a questionnaire the insured fills out, some external scanning, a binder, a policy. Controls are attested to. Premium is set. The carrier doesn't see the insured's security posture again until next year's renewal, unless there's a claim.

That model was defensible when the marginal cost of doing more was high and the threat landscape moved slowly. In a world where a new critical vulnerability can be weaponized in hours by a model running on commodity hardware, it is structurally inadequate. The underwriter is analyzing a posture that existed on the day the questionnaire was completed, against a threat that may not have existed yet.

You cannot underwrite cyber risk on annual snapshots anymore. You have to see the insured's attack surface continuously, verify controls continuously, and integrate threat intelligence continuously.

The unavoidable shift is from point-in-time, attestation-based underwriting to continuous, telemetry-based underwriting. You cannot underwrite cyber risk on annual snapshots anymore. You have to see the insured's attack surface continuously, verify controls continuously, and integrate threat intelligence continuously. Carriers that don't make this shift will find themselves adversely selected against by the ones that do. The carriers operating on continuous telemetry will know, in real time, which risks they should underwrite aggressively and which they should walk away from. The carriers still on annual questionnaires won't.

This is already happening. The hard market will arrive quietly, in the form of books of business bleeding margin to carriers with better signal, long before it announces itself as a pricing event.

Why Active Insurance was built for this

Coalition pioneered Active Insurance on the thesis that cyber risk is uniquely suited to a combined insurance-and-security product, because the data required to underwrite well and the data required to defend well are the same data. We don't sell a policy and wait. We continuously scan every policyholder's internet-facing attack surface. We ingest trillions of log lines from our customers' networks, both public-facing and, where they've agreed, private-facing. We run agentic underwriting infrastructure that reassesses exposure in near real time and reaches out to insureds when we see something concerning, often before they know themselves. When a vulnerability drops that we've concluded is material, we contact affected customers with patching guidance that day, not at the next renewal.

This matters in a Mythos-era world for specific, compounding reasons.

First, continuous telemetry means compressed weaponization windows don't catch us flat-footed. We don't wait for a claim to learn an insured was exposed. We see it in the data, and frequently before an attacker does.

Second, agentic underwriting means the marginal cost of continuous reassessment approaches zero. That is the only way an underwriting business model survives a world where the risk landscape moves hourly. Carriers doing the work manually cannot keep up. The unit economics don't work.

The carriers operating on continuous telemetry will know, in real time, which risks they should underwrite aggressively and which they should walk away from. The carriers still on annual questionnaires won't.

Third, a tech-first infrastructure means we can adopt defensive AI capability, including tooling analogous in posture to what Project Glasswing is doing for the largest software vendors, across our entire book of business. Not only the largest enterprise accounts. The small and mid-market insureds that traditional carriers increasingly struggle to underwrite profitably are exactly where automated, AI-assisted defensive capability has the biggest asymmetric return. This is precisely why we acquired Wirespeed, an Automated Detection & Response (ADR) platform that is built for speed and trusted by the Fortune 500.

Fourth, our claims data feeds directly into our underwriting models. The signal from a loss in one corner of the portfolio sharpens pricing across it within days, not at the next annual recalibration.

What this adds up to, in an era where the traditional carrier's annual questionnaire is becoming dangerously obsolete, is a structural advantage. Active Insurance isn't immune to systemic events. But it is the only cyber insurance model built from the ground up on the assumption that risk must be managed continuously rather than priced once a year.

What we honestly don't know

I want to resist over-claiming. The Mythos technical report is Anthropic's own evaluation, and while the UK AISI's independent work corroborates significant pieces of it, we are early in understanding how these capabilities generalize, how they degrade under real-world conditions outside a red team environment, and how quickly defensive tooling closes the gap.

There are legitimate open questions. How much of Mythos's reported capability reflects careful task scaffolding versus general model competence? How robust is autonomous exploit generation against targets hardened with runtime exploit mitigation, including Windows Defender Exploit Guard, macOS hardened runtime, and Android's seccomp-bpf filters? How quickly will defensive deployment by Project Glasswing participants, EDR vendors, and cloud providers narrow the offensive-advantage window?

Our working assumption is that defensive AI capability, if deployed aggressively and at scale, can close most of the window Mythos opens.

There is also a real question about proliferation. Anthropic has held Mythos back from broad release, and reports already indicate a small group of unauthorized users gained access the day of the announcement. How long does containment hold? What happens when an equivalent capability, which OpenAI has reportedly released to a limited group, is available through less disciplined channels?

Our working assumption is that defensive AI capability, if deployed aggressively and at scale, can close most of the window Mythos opens. That's the thesis behind Project Glasswing, and it's the thesis behind how we're upgrading our own defensive tooling. Aggressively and at scale is a large lift that will require the security and insurance industries to move faster than they historically have. We intend to.

In Closing

The Mythos moment is real. It is not the end of cyber insurability. It is, in our judgment, the end of point-in-time, attestation-based cyber underwriting as a viable business model, and the beginning of a period where the structural advantages of continuous-telemetry, tech-first insurance become impossible to ignore.

The sky is not falling. But the equilibrium that produced the current cyber insurance market is shifting, and it's shifting in a direction that rewards carriers who built their infrastructure for this kind of world and punishes those that didn't.

At Coalition, we've been quietly preparing for this moment for almost a decade. We think the next several years will make the case for Active Insurance more clearly than any pitch we could have written.

This communication represents the opinions and perspectives of the authors and is designed to provide general information on the topic presented. It is not intended to construe or render legal or other professional services of any kind. If legal or other professional advice is required, the services of a professional should be sought. Neither Coalition nor any of its employees make any warranty of any kind, express or implied, or assume any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, product or process disclosed. Any action you take upon the information contained herein is strictly at your own risk. Coalition and its affiliates will not be liable for any losses and damages in connection with your use or reliance upon the information. This communication may include links to other third-party websites. These links are provided as a convenience only. Coalition does not endorse, have control over nor assumes responsibility or liability for the content, privacy policy or practices of any such third-party websites. Coalition Insurance Solutions, Inc., an affiliate of Coalition, Inc., is a licensed insurance producer and surplus lines broker (Cal. license # 0L76155), acting on behalf of a number of unaffiliated insurance companies, and on an admitted basis through Coalition Insurance Company a licensed insurance underwriter (NAIC # 29530). See licenses and disclaimers. Copyright © 2026. All rights reserved. Coalition and the Coalition logo are trademarks of Coalition, Inc. All other products and company names are the intellectual property of their respective brand owners.