Anthropic just hit a major security snag that is every tech company’s worst nightmare. A group of unauthorized users reportedly gained access to Mythos, a highly restricted AI model designed specifically for cybersecurity. Anthropic built Mythos to help big companies protect themselves, but they kept it under tight lock and key. They feared that in the wrong hands, the model could become a dangerous weapon for hackers. Now, it looks like those exact “wrong hands” might have found a way in through the back door.

According to reports from Bloomberg, a group on a private online forum managed to get their hands on the tool through a third-party vendor. They didn’t just hack their way into Anthropic’s main servers. Instead, they exploited a weaker link in the chain. The group supposedly used the credentials of a third-party contractor who already had legal access to the Mythos Preview. This is a classic example of how even the most secure AI models are only as safe as the people who have the keys to the building.

Screenshots and Discord Drama

The group behind the leak isn’t staying quiet about it. They are reportedly part of a Discord channel that spends its time hunting for unreleased AI models. To prove they really have the goods, they shared screenshots and even gave a live demonstration of the software to Bloomberg. They claim they have been using Mythos regularly since they first gained access. For a company like Anthropic that prides itself on safety and “constitutional AI,” this is an embarrassing and potentially dangerous blow to their reputation.

Anthropic confirmed they are investigating the situation. A spokesperson said they are looking into reports of unauthorized access to the Claude Mythos Preview through one of their vendors. The company was quick to point out that their internal systems remain secure. They have found no evidence that the group’s activity has affected Anthropic’s own servers or data. While that might be true, the fact remains that a tool built for elite cyber defense is now being played with by a bunch of random people on the internet.

The Problem With Restricted AI

This leak highlights a massive problem in the AI industry. Companies like Anthropic create powerful models but refuse to release them to the public because they are “too dangerous.” They argue that keeping these tools restricted is the only way to ensure national security. But as we see here, restriction does not equal safety. When you limit access to just a few dozen organizations and contractors, you create a very valuable target for hackers and leakers.

If a group of people on a Discord server can get access to Mythos, you can bet that rival governments and professional cyber-criminals are trying even harder. Anthropic wanted to avoid a scenario where their AI helped people launch attacks. By failing to secure their third-party connections, they might have accidentally created the very situation they were trying to prevent.

The Fallout for Anthropic

This news comes right as Anthropic was starting to win over government agencies and big banks with their “security first” approach. If these clients think that Anthropic cannot even keep its own restricted models private, they might look elsewhere for their AI needs. Anthropic will likely have to overhaul how they handle third-party access and who they trust with their “frontier” models.

For the rest of us, this is a reminder that the AI world is still a bit like the Wild West. Even the smartest people in the room can lose control of their creations. As these tools get more powerful, the stakes for a single leaked password or a lazy contractor go through the roof. Anthropic’s Mythos was supposed to be a shield, but right now, it looks more like a liability.