Here’s what Anthropic and OpenAI are really doing: they’re changing how security vulnerabilities are discovered and analyzed at a fundamental, production level.
On April 7, Anthropic released a preview of its new Claude Mythos model and launched “Project Glasswing,” setting off a chain reaction in the market.
Mythos hadn’t even been publicly released before it sparked significant market anxiety. Where there’s a vulnerability, someone will find it—and someone else will weaponize it. The security landscape has never felt more precarious.
On the first trading day after the Mythos announcement, stocks for security firms like CrowdStrike, Palo Alto Networks, and Zscaler took a noticeable hit, with some dropping by double digits.
OpenAI, perhaps sensing the public’s desire for a technological counterpunch, made its own move.
On April 15, OpenAI unveiled GPT-5.4-Cyber: a “privileged model” for cybersecurity defenders. The company also announced an expansion of its “Trusted Access for Cybersecurity” (TAC) program, opening it up to thousands of vetted individual defenders and hundreds of teams responsible for critical software security.
On the surface, it looks like a race between two AI giants. But underneath, it’s a broader experiment by U.S. AI companies to redefine what security capability even means.
A veteran cybersecurity engineer admitted his first reaction to the news was a sense of crisis: “The key point isn’t how powerful the models are. It’s that the entire way we find vulnerabilities is about to change.”
Li Guanghui, founder and CEO of AI security firm BraneMatrix, put it this way: “Traditional cybersecurity problems are being algorithmically reconstructed. The security of the algorithms themselves will become AI-native security—the core technical capability of a new era.”
Why do these model releases create such a massive ripple effect? If vulnerability discovery can be scaled and replicated by a model, is the entire competency threshold of the cybersecurity industry being rewritten? How will competition change?
The Shift Is Here
Traditional cybersecurity has always relied on human experience, judgment, and hard-won accumulation.
“Someone finds a vulnerability, someone exploits it, someone patches it,” explained one senior security engineer. “Attack methods evolve, and defenses follow. So in the past, no one started with complete security capabilities, and no single company could stay ahead forever. Everything depended on accumulated knowledge.”
In this world, there’s always a lag between attack and defense. A vulnerability is discovered, exploited, and then patched. New tools emerge, and capabilities diffuse through practice. Whether it’s an internal corporate team or a vendor like CrowdStrike or Palo Alto Networks, the core task has been the same: productizing complex security capabilities and delivering them to enterprise clients.
But with major AI model companies entering the fray, a different path is emerging. Security capabilities are no longer solely dependent on people and experience, built through prolonged confrontation. Instead, some of that capability is being directly packaged into models, becoming a service you can call.
Whether it’s Mythos or GPT-5.4-Cyber, we need to be wary of the distribution logic and potential impact of these new forms of security power.
GPT-5.4-Cyber’s key feature is fine-tuning for defensive security use cases, lowering its refusal rate for legitimate security operations. Previously, if an engineer asked an AI to help test a system for vulnerabilities or analyze suspicious code, it might just say no—unable to distinguish defense from attack. GPT-5.4-Cyber is designed to solve that, “lowering the barrier” for security researchers. It also uses a tiered trust mechanism for deployment, which is more open than Mythos’s “allowlist” approach.
Anthropic’s Mythos wasn’t specifically trained for cybersecurity, but its general capabilities are so powerful they’ve “spilled over” into the security domain. It can understand code, discover vulnerabilities, and even reproduce and exploit them—reportedly outperforming previous models by a generation.
According to official disclosures, Mythos has already discovered thousands of zero-day vulnerabilities in mainstream operating systems and browsers, many classified as high-risk. In some capabilities, it’s said to surpass most security engineers, and the whole process requires minimal human involvement.
More important than the raw capability boost is this underlying reality: When a large model’s code comprehension and security analysis are strong enough, once restrictions are lifted or authorization is granted, it can be used to mine vulnerabilities and even convert them into attack capabilities. And when a model has this power, people are no longer the sole source of capability.
Disruption, Not Replacement
The consensus among multiple practitioners is clear: in the short term, this isn’t about replacement.
“People tend to oversimplify the problem,” said Li Guanghui. “Security capability isn’t a single technical skill; it’s an entire system.”
The traditional cybersecurity industry spans network protection, identity authentication, threat intelligence, risk control systems, and more. These areas can’t be covered by model inference alone.
Li believes that while the industry covers many niches, there’s still strong, fundamental demand for enterprise security services.
“Plus, traditional vendors have accumulated threat intelligence data—like non-public IOC (Indicators of Compromise) libraries containing IPs and file hashes (strings generated by specific algorithms to represent file content) related to black/gray market attacks. This remains a barrier that models like Mythos can’t easily overcome in the short term,” he said. “In scenarios like detecting black/gray market attacks or internet risk control that require specialized data, models don’t yet show a clear advantage.”
Furthermore, the “last mile” of security still depends heavily on people.
A senior penetration tester gave an example: A full penetration test often starts from an internet-exposed surface and works its way into a corporate intranet, even to core systems. This process involves navigating complex system environments and constantly bypassing non-standardized protections like CAPTCHAs, human verification, and dynamic policies. “Often, what determines success isn’t the tool itself, but on-the-spot judgment and adaptive strategy.”
Still, while full replacement isn’t imminent, disruption is inevitable.
The change is already happening. As large models rapidly take over highly manual tasks like vulnerability analysis, code auditing, and attack path simulation—operating at lower cost and higher frequency—the value structure of the security industry is being reshaped: foundational capabilities are being compressed, while higher-order skills are being amplified.
“Once this shift occurs, the entire industry’s operating logic will be reconstructed,” said a veteran security researcher. “As AI’s coding and comprehension abilities improve, security capability itself will be massively amplified. The traditional system relying on human experience will gradually weaken. This isn’t replacement; it’s a shift in the industry’s center of gravity.”
The Risks and the Real Crisis
Anthropic states that Mythos is a “defense-only” model, open only within a closed alliance of 52 tech companies, including Amazon, Apple, Broadcom, Cisco, Google, Microsoft, and Nvidia.
The company’s rationale is that Mythos’s safety guardrails aren’t mature yet. A full release would hand attack tools to both defenders and attackers. So, the model maker wants partners to use it first, patch critical infrastructure vulnerabilities, and then gradually open access.
That reasoning makes sense, but it’s also causing concern in the industry. Because in real-world security, “defense” is never a pure concept. Practices like penetration testing, red team exercises, and vulnerability validation—which look like “attacks”—are actually the foundation of defense.
For instance, research has shown that while AI companies place many restrictions on their advanced models to prevent misuse, some models can still bypass these limits through unconventional methods (like using classical Chinese, ancient Sanskrit, or Latin as “circumlocution”). In specific scenarios, they might enter a state with almost no output constraints. So, if someone can bypass these restrictions, they could use these models for attack, not just defense.
In other words, defense is essentially “attack simulation for the purpose of protection.” When a model’s code comprehension and security analysis are strong enough, the line between offense and defense blurs rapidly—and it becomes very hard for a vendor to precisely control that boundary.
Beyond security, there are other interpretations.
Venture capitalist Marc Andreessen questioned whether Anthropic is truly restricting release due to safety concerns or because it lacks sufficient computing power to support large-scale public use. He noted that Anthropic has recently experienced frequent service outages and has limited user compute during peak hours.
Li Guanghui disagrees with Anthropic’s approach: “This so-called ‘defensive nature’ essentially prioritizes model access for U.S. software and cloud service vendors like Apple, Google, Microsoft, and IBM. Security should be validated through open confrontation, not unilaterally defined and endorsed by a few entities. If the standards, test environments, and results of security evaluation aren’t transparent, then this ‘security’ itself lacks credibility.”
Here’s the crucial point: whether it’s Anthropic’s Mythos or OpenAI’s GPT-5.4-Cyber, these are security tools being prioritized for the U.S. tech ecosystem.
When a key technology is made available only to specific countries or a handful of institutions—rather than diffusing across the global industry—it tends to create “technical barriers” and even “technical dominance.”
We’ve seen this pattern before in other domains: from long-term U.S. export restrictions on NVIDIA’s high-end GPUs (like the H200 and B300) to early controls on encryption technology. When core capabilities have restricted flow, industry structures inevitably change.
This deviates from a long-standing practice in cybersecurity. Over decades, “ethical hacking” became mainstream, paired with responsible coordinated disclosure mechanisms that encouraged more security researchers to participate. By finding and disclosing vulnerabilities, they helped improve overall system security. While not a rule everyone follows perfectly, it has been a foundational principle for how the traditional cybersecurity industry operates.
We need to ask: in a closed alliance like this, will the power to define security begin to concentrate in the hands of a few countries or entities? If so, the future of cybersecurity assessment, verification paths, and even who gets to speak authoritatively about it could be up for grabs.
All these changes lead back to a more practical question: how will every participant in this industry reposition themselves?
For traditional security vendors, the path forward needs to be figured out—and soon.