By Karthik Ramakrishnan
Anthropic's announcement of Claude Mythos has landed at the intersection of frontier AI capability, cyber risk, and insurance pricing.
The headline claim is dramatic. Anthropic says Mythos can autonomously identify and, in some cases, exploit thousands of serious software vulnerabilities across major operating systems, browsers, and other critical infrastructure. Some of those bugs had reportedly been dormant for decades. Anthropic has chosen not to release the model publicly and has instead restricted access through Project Glasswing, a coalition that includes AWS, Apple, Microsoft, Google, CrowdStrike, Palo Alto Networks, NVIDIA, JPMorgan Chase, Cisco, Broadcom, and the Linux Foundation.
That framing has produced two immediate reactions. One camp sees Mythos as a genuine inflection point in AI enabled cyber capability. The other sees a familiar pattern of frontier labs wrapping product strategy in safety theater.
As someone who underwrites AI risk, my view is more measured. Mythos looks real enough to matter, but not proven enough to justify the most apocalyptic claims. For insurers, that is already enough.
There are several reasons not to dismiss this as marketing.
First, some of the technical details appear to line up with real world remediation. OpenBSD's March 25, 2026 errata confirms a kernel patch for invalid TCP SACK options, matching Anthropic's claimed 27 year old discovery. CVE-2026-4747, a FreeBSD vulnerability attributed to Mythos, appears in the National Vulnerability Database. Those datapoints do not prove every headline claim, but they do suggest a real step forward in vulnerability discovery and exploit reasoning.
Second, the surrounding coalition matters. CrowdStrike and Palo Alto Networks did not build their businesses by lending credibility lightly. Nor did JPMorgan Chase, Microsoft, Google, or Apple. Their participation does not mean every public claim is validated. It does mean serious institutions appear to believe there is something here worth treating seriously.
Third, the government response suggests the issue is not being treated as a press cycle artifact. Reports indicate that US Treasury Secretary Bessent convened a meeting with the chief executives of major American banks to discuss Mythos. Canadian bank executives and regulators reportedly held a similar emergency session. Those are signals of concern, even if the public still has only a partial view of the underlying briefings.
Finally, Anthropic has put forward at least one falsifiable metric. The company says expert reviewers agreed with the model's severity assessment 89 percent of the time across 198 manually reviewed reports, and were within one severity level 98 percent of the time. That is still a sample, not universal proof, but it is more substantial than pure narrative.
The skeptical case is not frivolous.
Chamath Palihapitiya argued on the April 10, 2026 All In Podcast that the Mythos rollout looks at least partly like commercial theater. David Sacks made a related point, noting Anthropic's pattern of pairing alarming safety disclosures with major launches. That criticism deserves to be taken seriously. AI labs have earned skepticism when they frame releases through the language of exceptional danger.
There is also a meaningful technical distinction between finding a valid vulnerability and finding one that is easily weaponized at scale. Anthropic itself acknowledged that at least one showcased example, an FFmpeg vulnerability, is likely not critical severity and may be difficult to weaponize. That matters because insurance loss trends are driven by exploitability, propagation, and concentration, not by bug counts alone.
The developer community has been digging further into the technical details. Tom's Hardware noted that Mythos found several potential exploits in the Linux kernel but was unable to actually exploit any of them because of Linux's defense in depth security systems. In its OSS-Fuzz-style testing of over 7,000 open source software stacks, Mythos found crashable exploits in around 600 and only 10 severe vulnerabilities. That is a meaningful improvement over prior models, but it is a long way from thousands of devastating exploits. The gap between what Anthropic's framing implies and what the granular results show is worth noting.
The other reason for caution is that Mythos may not be unique for long, or even now. Researchers at AISLE have argued that some of the vulnerabilities Anthropic highlighted may already be detectable by smaller or more open models once the relevant code paths are isolated. If that view is right, then the real story is not that Anthropic alone has crossed a threshold. It is that the broader market may be approaching one.
That would make the underwriting implications more important, not less.
The most useful way to think about Mythos is not as a binary question of true or false. It is a scenario problem.
Even the first scenario should get the market's attention. The second and third require action.
This is where Mythos matters most.
The cyber insurance market entered this year in a relatively comfortable position. Direct written premiums declined 2.3 percent in 2024, the first decrease since data collection began in 2015. Capacity has been plentiful. Competition has been intense. Many carriers have felt more confident as policyholder hygiene improved and recent loss ratios remained manageable.
Mythos puts pressure on the assumptions underneath that comfort.
The core issue is not simply attack frequency. It is accumulation risk and correlation.
Cyber insurance has long relied on an implicit premise that serious zero day discovery and exploitation are scarce, expensive, and concentrated among highly capable actors. If AI compresses that process, then the exposure profile changes. Risk shifts from isolated incidents to simultaneous weakness across portfolios built on the same operating systems, browsers, cloud platforms, identity infrastructure, hypervisors, and open source components.
That is the underwriting shock embedded in the Mythos story.
A latent flaw in a common dependency can move from obscure technical debt to portfolio wide event much faster than most current exposure models assume. That is the scenario that should concern carriers and reinsurers most. Not whether one company made a dramatic announcement, but whether AI is increasing the speed at which hidden technical debt becomes correlated insured loss.
A few implications follow directly:
The implications for technology errors and omissions insurers are distinct, and in some ways more legally consequential.
If AI can systematically surface vulnerabilities in commercial software that remained undetected for years, the unknown defect defense weakens. Plaintiffs will increasingly argue that flaws were discoverable with available tools and that reasonable diligence should have identified them earlier.
That affects several classes of insureds:
This is where cyber and tech E&O begin to converge more visibly. Silent cyber concerns, wording overlap, trigger disputes, and questions about whether a loss sits in product liability, professional liability, or cyber coverage will all become more important.
The wrong response is to dismiss Mythos as marketing and wait for clean loss data.
The other wrong response is to accept the most dramatic framing at face value and underwrite as though the internet has already changed overnight.
The prudent stance is more disciplined.
My base case is that Mythos is materially real, but directionally over marketed. Anthropic is probably pointing at a genuine inflection in AI enabled cyber capability, even if the most expansive claims around scale and exclusivity prove exaggerated. For insurers, that is enough to act on.
That means:
This is not a call for panic. It is a call for underwriting discipline.
That is the conclusion I would draw from Mythos.
Not that the internet is over.
But that the assumptions underneath cyber and tech E&O underwriting are being pressured by AI sooner than many in the market, and many carriers, ever expected.
Karthik Ramakrishnan is the Founder and CEO of Armilla AI, a Lloyd's of London Coverholder and MGA specializing in AI insurance and liability coverage.
Disclaimer: The insurance products described on this website are offered through surplus lines insurers. Surplus lines insurance is not subject to all the insurance laws and regulations of your state, and such insurers are generally not licensed by or subject to the supervision of the insurance department of your state. Coverage may not be available in all jurisdictions and is offered only through properly licensed surplus lines brokers.
Nothing on this website constitutes an offer to sell or a solicitation to buy insurance. No coverage will be bound or effective unless and until confirmed in writing by an authorized representative and all underwriting requirements have been satisfied. Any descriptions of coverage are for general informational purposes only and do not amend, modify, or supplement any actual policy or contract of insurance.
Please consult with a licensed surplus lines broker to determine availability and eligibility in your state.