Anthropic said this week it is holding back the public release of Mythos, its newest large‑language model, because the system can locate security flaws in software that underpins global online services. The company will instead make the model available to a handful of major firms that operate critical internet infrastructure, such as Amazon Web Services and JPMorgan Chase.

The decision reflects a growing tension in the AI community between rapid innovation and the responsibility to safeguard the digital ecosystem. Anthropic argues that unleashing a model capable of surfacing zero‑day exploits could give malicious actors a powerful new tool, potentially accelerating attacks on widely used platforms.

Industry observers note that Anthropic’s caution aligns with a parallel strategy aimed at bolstering enterprise revenue. By limiting Mythos to high‑value customers, the frontier lab creates a “flywheel” that locks in large contracts while making it harder for smaller labs to replicate its capabilities through model distillation. Distillation—a process that uses a powerful model to train a smaller, cheaper one—has emerged as a threat to the business model of firms that invest heavily in scaling compute.

David Crawshaw, CEO of the startup exe.dev, framed the move as marketing cover for a deeper financial motive. He warned that by the time open‑source researchers gain access to Mythos, Anthropic will likely have launched a newer, enterprise‑only version, further cementing its dominance in the high‑end market.

Anthropic’s stance also mirrors actions taken by OpenAI, which is reportedly contemplating a similar restricted rollout for its upcoming cybersecurity tool. The companies appear to be coordinating efforts to identify and block parties that attempt to copy their models, especially firms based in China, according to a Bloomberg report.

Critics, however, argue that the security rationale may be a convenient pretext. Dan Lahav, CEO of AI cybersecurity lab Irregular, emphasized that the real danger lies not just in finding a vulnerability but in whether that flaw can be exploited in a meaningful way. He questioned whether Mythos’ discoveries translate into actionable attacks or remain academic curiosities.

Competing AI security startup Aisle pushed back against the notion that Mythos is a singular solution. The firm demonstrated that many of the model’s claimed achievements could be reproduced with smaller, open‑weight models, suggesting that no single LLM can dominate the entire cybersecurity landscape.

Anthropic has not confirmed whether the limited release also serves to impede distillation efforts, but the timing dovetails with a broader crackdown on model copying. The company recently publicized alleged attempts by Chinese firms to duplicate its technology, and it has joined forces with Google and OpenAI to track and block such activities.

Whether Mythos will prove a decisive advantage in defending the internet remains uncertain. The cautious rollout could allow large enterprises to pre‑empt potential threats while giving Anthropic a competitive edge in the lucrative enterprise AI market.

For now, the AI community watches closely as frontier labs balance the promise of powerful new models against the risks of exposing them too broadly.

Used: News Factory APP - news discovery and automation - ChatGPT for Business