Limited release of Anthropic’s Claude Mythos raises AI access concerns

In short: A recent security scare around Anthropic’s Claude Mythos is also drawing attention to a growing issue, not everyone can get access to the most powerful AI models.

What's going on

Anthropic released a new AI model called Claude Mythos to only a small number of technology customers. The idea was to help them find and fix security flaws in widely used software. A security flaw is like a weak lock in a door, it can let attackers in.

The model raised alarm because it may be able to spot serious holes in software that many companies rely on, including banks. Anthropic did not fully share the test results behind its warning, which makes it hard for outside researchers to confirm the claims. Security expert Bruce Schneier told the Financial Times the warning could have happened months earlier or later, but the broader risk from AI-assisted hacking is real.

At the same time, limited access has an economic effect. Schneier said Anthropic is already struggling to meet demand and may not have enough computing power to share Mythos more widely. Computing power is the large amount of machine time needed to run these models, like having too few engines to power too many trains.

What to watch

OpenAI reportedly released a similar model to a limited group of customers this week, which suggests restrictions could become more common. If only some companies can say their products were checked by a top model, like being “Mythos-vetted,” they may gain an advantage over rivals. Watch for whether AI companies limit access mainly for safety reasons, or because scarce supply lets them charge more and choose the most profitable customers.

Source: Financial Times