NextFin News - Anthropic PBC has expanded access to its most advanced artificial intelligence model, Mythos, granting 150 additional organizations entry to a research preview previously restricted to a handful of elite partners. The move, announced on June 2, 2026, marks a significant widening of "Project Glasswing," a defensive cybersecurity initiative designed to test the model’s capabilities in identifying and patching critical software vulnerabilities before they can be exploited by malicious actors.
The expansion comes just one day after Anthropic filed confidentially for an initial public offering (IPO), signaling a strategic push to demonstrate the commercial and safety-oriented utility of its "frontier-class" technology. Mythos, which Anthropic internal documents describe as a "tier above" its current flagship Opus models, has been kept under tight wraps due to concerns over its potential misuse in offensive cyber operations. By increasing the pool of testers from an initial 12 organizations to over 160, the company is attempting to balance the rapid scaling of its technology with the rigorous safety protocols that have become its brand hallmark.
The decision to broaden access is viewed by some as a response to the surging demand for high-reasoning AI in the enterprise sector. "Anthropic is under immense pressure to prove that its safety-first approach doesn't come at the cost of market share," said Marcus Thorne, a senior technology analyst at Arcus Research. Thorne, who has historically maintained a cautious stance on the rapid deployment of large language models, noted that this expansion is likely a "controlled stress test" intended to satisfy both potential IPO investors and federal regulators who are increasingly wary of AI-enabled cyber threats.
Thorne’s perspective reflects a broader debate within the industry regarding the "dual-use" nature of models like Mythos. While Anthropic emphasizes the defensive benefits—citing a recent partnership with Mozilla that saw Firefox security bug fixes jump from double digits to over 400 in a single month—the risk of "model leakage" remains a persistent concern. Reports surfaced last week of unauthorized users briefly gaining access to Mythos-tier capabilities through a misconfigured API endpoint, a lapse that Anthropic has since addressed but which underscores the volatility of managing such powerful assets.
The 150 new groups include a mix of cybersecurity firms, academic research labs, and infrastructure providers. Unlike the standard Claude API, which is available to the general public, Mythos remains invitation-only with no self-serve sign-up. This tiered approach allows Anthropic to maintain a "kill switch" over the model’s deployment while gathering the diverse datasets necessary to refine its reasoning capabilities. However, this exclusivity has drawn criticism from smaller developers who argue that the concentration of such powerful tools among a few "trusted" entities could stifle innovation and create a new class of digital divide.
From a financial standpoint, the timing of the Mythos expansion is inextricably linked to Anthropic’s public market ambitions. With co-founders Dario and Daniela Amodei now overseeing a company valued at over $40 billion following its latest funding round, the pressure to convert research breakthroughs into defensible revenue streams is at an all-time high. The success of Project Glasswing will likely serve as a primary case study for the company’s ability to monetize high-stakes AI applications in regulated industries like finance and national defense.
The broader market remains divided on whether Anthropic can maintain this restrictive distribution model indefinitely. While the company has stated it does not currently plan a general release for Mythos, the history of the AI industry suggests that competitive pressures from OpenAI and Google often force a transition from "research preview" to "commercial product" faster than originally intended. For now, the 150 additional groups represent a middle ground—a way for Anthropic to scale its influence without fully relinquishing control over its most potent intellectual property.
Explore more exclusive insights at nextfin.ai.
