Anthropic unveils Mythos, an ultra-secure and expensive AI model

Anthropic Prepares to Launch Mythos

Anthropic, a company specializing in artificial intelligence, recently announced significant advancements regarding its AI model, Mythos. Currently, this model is reserved for a consortium of partners as part of the Project Glasswing. The company plans to make it available to the general public in the coming weeks, after strengthening the necessary protections. According to Anthropic, Mythos possesses advanced capabilities in cybersecurity, justifying its limited access to experts to address vulnerabilities before a wider release. The company claims to be making rapid progress in developing the necessary cybersecurity protections for widespread use.

Anthropic's decision to withhold Mythos Preview from the public, at least for now, is intriguing. It raises questions about the protection of critical Internet infrastructure against potential vulnerabilities. Some wonder if this decision is a precautionary measure or a marketing strategy to generate buzz. Security researchers have found that the model can detect exploits much faster than human hackers, although it does not necessarily surpass human capabilities. For example, the latest version of Mozilla's Firefox included over 200 fixes identified by Mythos Preview.

Darren Williams, founder and CEO of cybersecurity firm BlackFog, stated that major model releases are often tense moments. "On one hand, Anthropic's decision to plan the release while waiting for protections to be developed shows good instincts," he said. "But the more capable a model is, the higher the stakes if those protections fail or if the model is ultimately misused. The period between the release of a powerful model and the widespread adoption of defenses against it is always a vulnerable time."

An Expensive Yet Promising Model

The Mythos model, while effective, comes with a high operational cost, approximately 30 times higher than previous models like Opus. Jake Williams, a cybersecurity researcher and faculty member at IANS Research, pointed out that this cost could limit access for convenience hackers, although state actors already have advanced technologies. "It's out of reach for many, including convenience threat actors," Williams said. "State actors already had better technology to find vulnerabilities. This only changes the game for a small percentage of threat actors."

Improvements with Claude Opus 4.8

Meanwhile, Anthropic has launched an update to its Claude model, with version Opus 4.8. This version features notable improvements over Opus 4.7, including a "modest but tangible improvement," according to Anthropic. Tests revealed that Opus 4.8 is less likely to make unsupported claims and more inclined to indicate uncertainty. New features also allow for controlling the "effort" the model expends in responding to a query, thereby influencing the quality and speed of responses. A higher effort will likely yield better results, as the model will spend more time on a response, but this will quickly deplete your usage limits. A lower setting will respond more quickly and reach rate limits more slowly.

These developments reflect Anthropic's commitment to enhancing its AI models while ensuring the security and efficiency of their applications.