Shaking Pandora’s Box: How Claude Mythos Upends the World of Cybersecurity

Defining Mythos

Anthropic’s Claude Mythos Preview is the latest advancement in AI models for autonomous coding and reasoning. What makes Mythos different from earlier models is its ability to search for weaknesses in existing software systems.

In the past, companies would invite outside security researchers to probe their systems and pay them for each verified bug they uncovered. Mythos upends this arrangement by shrinking the time required to identify vulnerabilities and by finding them at scale. This quickly exceeds the capacity of existing defensive measures. For instance, Cloudflare, an essential security gatekeeper for over 24 million active websites, discovered 2000 vulnerabilities using Mythos Preview within a month. In contrast, Cloudflare’s well-established and popular bug bounty program reportedly only patched 359 vulnerabilities since it was launched in 2014, at a cost of over $500,000 in bounties paid to hackers.

Essentially, Mythos has revealed how AI will disrupt the ability of organizations to keep up with software vulnerabilities. We believe that the limiting factor in cybersecurity will no longer be the ability to discover vulnerabilities but rather the capacity to fix them.

Sending Signals

Anthropic’s restraint is a signal: the company that built Mythos has decided it is too dangerous for unrestricted release. Their caution is not paranoia. AI‑generated cyberattacks are no longer theoretical. They are happening now, and they expose a truth that has been too easy to ignore: the world’s digital infrastructure is built on foundations that are far more fragile than most people realize.

Not everyone agrees with Anthropic’s decision to withhold Mythos from public release. Critics cite a long-standing principle in the cybersecurity field: security through obscurity is a dangerous and bad idea. By restricting Mythos to a few vetted organizations, a gap will widen between those who can find vulnerabilities and those who can fix them. Their view is that Mythos should be available to defenders broadly, not just to a select group.

Accumulating Debt

Technical debt refers to the accumulated consequences of shortcuts, legacy decisions, outdated technology, incomplete documentation and general lack of interest in maintaining old systems rather than building new ones. It has been a worry for organizations since the first computer-based systems emerged in the 1960s.

Many systems that underpin modern life such as banking platforms, airline reservation systems, hospital networks and municipal infrastructure were built decades ago. They have been patched, extended, and wrapped in layers of middleware, but rarely redesigned or replaced by modern systems. The people who originally built them have long since left their organizations.

From a cybersecurity standpoint, technical debt provides an enticing opportunity and a rich attack surface for exploitation. In the past, there was a sort of equilibrium between cybersecurity offense and defense. Hackers would develop new malware, and organizations would concurrently develop defenses to counter it. With Mythos and similar AI-based systems on the horizon, this long‑standing status quo will be significantly disrupted.

On Offense

In May, Anthropic announced that Project Glasswing, their initiative to allow approximately 50 companies access to Mythos, had led to more than 10,000 high- or critical-severity vulnerabilities being discovered just since going live last month. This showed how Mythos was valuable for defense. But there is little doubt that AI-based systems can be used for offensive purposes as well.

Google’s Threat Intelligence Group recently reported that attackers used an AI system to: (1) identify a novel vulnerability; (2) generate a working exploit; and (3) prepare a large‑scale attack. This is the first confirmed case of an AI system autonomously generating a zero‑day exploit. Google did not disclose the intended targets but emphasized that this incident is likely “the leading edge” of a broader trend.

Machine Speed

The unfortunate answer is ‘they are not the same’.

First, discoveries need human verification. Mythos surfaces vulnerabilities at machine speed but confirming them still requires human assessment. This is slow, judgment-heavy work, even with AI assistance.

Second, patching is fundamentally harder than finding. It requires understanding legacy code, designing a safe patch, testing the patch across different environments without introducing new vulnerabilities, ensuring the remediated software works effectively, and coordinating releases.

Third, the institutional environment is the real bottleneck. Disclosure norms, legal review, vendor coordination, and customer-impact assessments all add complications to any changes. Even though AI accelerates discovery, institutional practices do not keep pace. It is easier for a lone hacker to adopt and use AI systems, whereas organizations need institutional changes to adopt and implement AI systems.

Finally, in many systems, the vulnerability is not a bug but a design decision. Patching requires re-architecting, not editing. AI can point out the crack, but it can't rebuild the bridge. The result is a structural imbalance: AI has collapsed the cost of knowing, but the cost of fixing remains stubbornly high, time-consuming, and a major bottleneck.

New Solutions

We can see a number of strategies to help deal with the challenges posed by this new reality.

AI-Assisted Remediation

The best‑case scenario is that AI systems begin to help organizations fix vulnerabilities as quickly as they are discovered. They already can produce code at scale. Now this ability needs to be channeled into remediation.

Harden Systems

Some front-facing components of systems, such as websites or APIs, are easy for AI to access. Organizations should harden them by focusing on stronger network protocols, enforcing multi-factor authentication, and improving employees’ security compliance.

New Institutions & Regulations

We may see the rise of centralized remediation platforms or “cyber-FEMA” ‑ style organizations. Governments may also need to impose stricter requirements around disclosure and patch timelines since unpatched vulnerabilities could lead to significant societal disruption. Hospitals, water treatment plants, power grids and other important utilities or services all run the risk of being compromised by AI-exploited vulnerabilities.

Legacy Systems Must Be Replaced

For far too long, organizations have shied away from replacing long-standing legacy systems. They fear the cost and effort, or do not see a need to replace something that still technically works. Unfortunately, with AI-based systems like Mythos in the hands of bad actors, organizations will no longer be able to delay or hesitate. Whether they like it or not, they will either have to rebuild critical systems from scratch or they will have to buy/rent software from trusted vendors. Another alternative, which some companies are exploring, is to use sophisticated AI-based software development systems to rewrite the legacy systems’ code in-place. However, this approach brings its own challenges and may introduce new weaknesses even as old ones are alleviated.