Experts warn that the threats highlighted by Mythos are not new and have been achievable with existing AI models for some time.
Key Points: - Cybersecurity experts and AI researchers confirmed to CNBC that the high-risk capabilities demonstrated by Anthropic's Mythos model can already be replicated using older, publicly available AI models from companies like Anthropic and OpenAI. - While AI accelerates the discovery of software vulnerabilities, organizations often require days or weeks to develop and deploy patches, creating a widening security gap and leaving systems exposed. - Researchers note that in the initial phase of the AI era in cybersecurity, the inherent advantage lies with attackers, not defenders, despite ongoing efforts by companies like Anthropic and OpenAI to develop defensive capabilities. - In response to CNBC, Anthropic did not deny that its earlier models possessed the ability to discover software vulnerabilities.
On February 19, 2026, Dario Amodei, co-founder and CEO of Anthropic, attended the AI Impact Summit in New Delhi, India. This week, India hosted one of the world's largest AI summits as the Modi government aims to position the country as a hub in the competitive landscape of cutting-edge AI models.
Last month, major global banks, tech giants, and governments scrambled to mitigate risks associated with Anthropic's new model, Mythos. The powerful model had already uncovered thousands of previously unknown security vulnerabilities within global software infrastructure.
However, the critical issue is that the threat causing widespread panic today has, in fact, existed for some time.
Cybersecurity professionals and AI researchers told CNBC that the types of software vulnerabilities Mythos can find are already within the reach of models available on the market, including older versions from Anthropic and OpenAI.
Ben Harris, CEO of cybersecurity firm watchTowr Labs, stated, "The entire industry can already reproduce similar vulnerabilities to those found by Mythos by rationally orchestrating existing public models, with highly similar end results."
The release of Mythos has caused significant anxiety among corporate executives and policymakers, fueling fears of a new era of AI-powered cybercrime. To reduce the risk of the model falling into the wrong hands, Anthropic has limited its release to a select group of U.S. companies, including Apple, Amazon, JPMorgan Chase, and Palo Alto Networks.
Despite this precautionary measure, the Trump administration is reportedly considering new government regulations for future AI models as a result.
The high-profile launch of Mythos represents the latest step in a series of prominent product updates from Anthropic, intensifying its competition with OpenAI. Both AI giants are preparing for highly anticipated IPOs. Weeks after Mythos's release, OpenAI CEO Sam Altman unveiled GPT-5.5-Cyber—a specialized model tailored for cybersecurity.
OpenAI made GPT-5.5-Cyber available on a limited basis to vetted cybersecurity teams this past Thursday.
The controlled, limited release of Mythos is part of a security initiative called the "Glasswing Project." Its intent is to give organizations time to bolster their cyber defenses ahead of potential large-scale attacks from criminal organizations or hostile nation-states.
Anthropic CEO Dario Amodei said at a company event this week, "The real risk is a significant surge in the number of software vulnerabilities, cyber intrusions, and financial losses from ransomware, with schools and hospitals being the first targets, not to mention major banks."
The Already Present Threat Industry professionals on the front lines of cybersecurity state that the core capability heavily promoted by Anthropic—the large-scale, batch discovery of software vulnerabilities—has been achievable since last year.
Claudia Cloch, CEO of cybersecurity firm Vidoc, told CNBC, "The models we are currently using are already sufficient for detecting zero-day vulnerabilities at scale, which is alarming in itself."
She noted that this capability has existed for at least several months, possibly up to a year.
A zero-day vulnerability is a software flaw that is not publicly known and for which no official patch exists, allowing attackers to exploit it before defenders can respond.
Vidoc's research team employed a model orchestration technique to test whether they could replicate Mythos's vulnerability discovery results. This method involves breaking down code into modules and using multiple tools and AI models to cross-verify findings.
Cloch said, "We tested the same codebase with older models, and both OpenAI's and Anthropic's earlier models successfully identified the same types of vulnerabilities."
Another cybersecurity firm, Aisle, also found that many of Mythos's signature achievements could be replicated by orchestrating multiple lower-cost models in parallel. This suggests that model orchestration and synergy are more critical than using the very latest version.
Stanislav Fort, founder of Aisle, wrote in a blog post, "A thousand decent detectives conducting a comprehensive search will find far more program flaws than a single top detective guessing where to look."
In its response to CNBC, Anthropic did not deny that its earlier models had the capability to discover software vulnerabilities.
A company spokesperson stated that Anthropic has been warning for months about the rapid evolution of AI capabilities in cyber offense and defense. The company published a blog post in February disclosing that the publicly available Claude Opus 4.6 model had already found over 500 high-severity vulnerabilities in open-source software.
Amodei confirmed this point again at this week's event: while Mythos discovers vulnerabilities at a scale far surpassing previous models, this trend in capability evolution has been evident for some time.
"The risk is real, which is why we implemented a limited release. But to some extent, this is not surprising; we have seen these risk warnings for a while."
Widespread Collective Panic An Anthropic spokesperson stated that Mythos's unique aspect is its ability to autonomously generate exploit code for vulnerabilities with minimal human intervention, automating a process that previously required senior researchers.
However, cybersecurity researchers point out that criminal groups and state-sponsored hackers from adversarial nations have long possessed such techniques. Cloch stated bluntly, "Hackers from North Korea, Russia, and others are fully capable of doing this even without the Anthropic model."
Harris noted that the potential threat of AI-empowered hacking has put companies and regulators on high alert, fearing a new wave of ransomware and various cyberattacks targeting critical infrastructure. He described the communication atmosphere with banks, insurance firms, and regulators in recent weeks as one of "collective panic."
Even before the rise of generative AI, the industry faced a structural dilemma: skilled hackers could exploit newly discovered vulnerabilities within hours, while organizations often needed days or weeks to develop and deploy patches. Applying patches to some critical systems may require taking them offline, further complicating operations.
Harris said, "The industry is already overwhelmed and anxious about the sheer volume of security vulnerabilities. Even without widespread access to Mythos, organizations' patching speeds cannot keep up with the pace of risk emergence."
Previously, only a handful of specialists globally had the capability and time to discover vulnerabilities in niche software and launch attacks. Now, with readily available AI models, the barrier to launching large-scale cyber disruption has significantly lowered.
This means core targets like banks will face more attacks, and software systems previously overlooked by cybercriminals are now under threat.
The Offense-Defense Landscape: Inherent Advantage for Attackers Researchers believe that although companies like Anthropic and OpenAI are simultaneously developing cyber defense technologies to match the risks, in the cyber offense-defense dynamic, the initial advantage naturally tilts toward the attacker.
JPMorgan Chase CEO Jamie Dimon expressed a similar view last month: while AI tools may help strengthen cyber defenses in the long term, in the short term, they are likely to increase system vulnerabilities.
Justin Herring, a partner at law firm O'Melveny & Myers and former Deputy Superintendent for Cybersecurity at the New York State Department of Financial Services, stated, "We are seeing an explosion in the number of vulnerabilities being discovered, but the industry has not seen a corresponding emergence of efficient remediation tools. Vulnerability management has become a Sisyphean task in cybersecurity."
While the first institutions granted access to Mythos gained a head start in patching vulnerabilities, a downside emerged: independent AI researchers cannot access the model. This prevents them from verifying Anthropic's claims or proactively building targeted defenses.
The industry views this closed approach as hindering collective risk response across the cybersecurity sector.
Pavel Gurvich, CEO of cybersecurity startup Tenzai, stated that this creates a tiered division of security resources, widening the gap between institutions that can access such models and ordinary companies, potentially slowing innovation across the cybersecurity industry.
Ben Seri, co-founder of cybersecurity startup Zafran Security, said, "The industry hopes to find the optimal global security governance solution before such high-risk AI capabilities become widespread. It's a classic chicken-and-egg dilemma, and there will inevitably be costs along the way that cannot be avoided."
Comments