Anthropic could be close to launching new AI models with Mythos-class cybersecurity capabilities to the general public, less than two months after warning that those capabilities could exploit zero-day vulnerabilities and allow non-experts to produce sophisticated exploits.
It has reportedly built up considerably stronger safeguards in that time, to the point that it may be able to offer some of Mythos’ capabilities to a general audience. Anthropic had been debating whether to ever give Mythos a wider release, with some inside the company concerned about the negative effects it could have in the hands of bad actors.
The company made the admission alongside the release of Opus 4.8, an upgrade to its current most powerful model available to a general audience. It is not clear whether Anthropic would use the Mythos moniker if or when it releases the model.
From global security threat to general availability
The AI research lab has given Mythos Preview access to a select group of security teams, software maintainers, infrastructure companies, and governments under Project Glasswing. The aim was to give them time to shore up their defenses and identify vulnerabilities before rogue AI catches up.
According to Anthropic, it is about six to 18 months ahead of open-source AI models.
This head start has been welcomed by some in the industry. Mozilla reported that its latest release of Firefox included 271 vulnerabilities identified by Mythos. Anthropic also reportedly identified more than 23,000 vulnerabilities in open-source projects, with 6,200 estimated to be high or critical severity.
The potential jump in sophistication has also led many to try to gain access to the AI model, among them European finance and government leaders. The Trump Administration, while publicly feuding with Anthropic over its Pentagon contract, has given several departments access to Mythos, including the National Security Agency.
Anthropic may see opening up the AI model as necessary to ensure a level playing field, with developers as able to spot vulnerabilities as attackers. At least one unauthorized group gained access to Mythos Preview, according to a report last month, and the clock is ticking on how long Anthropic can maintain its advantage in cyber capabilities.
Positioning before the IPO bonanza
Its main rival, OpenAI, has been more freely distributing its own cyber model to other countries and large organizations. OpenAI CEO Sam Altman called out Anthropic's warnings as “fear-based marketing,” as the two companies have differed on the right way to distribute their models.
This split could become one of the key marketing messages from both camps before their IPOs. Anthropic has signaled that it is more risk-averse, with highly sophisticated and powerful models that can be used for good. On the other side, OpenAI has shown a willingness to put its own powerful models in the hands of more people, ultimately judging its safeguards to be adequate to stop bad actors.
Anthropic is on the ascendancy at the moment, having recently closed a funding round at a higher valuation than its rival, OpenAI. This is partly due to a shift in investor interest and valuation across the AI market, with enterprise customers becoming much more valuable than users. OpenAI still leads the way with 900 million active users, but Anthropic has reportedly matched its revenue run rate and has more business customers.
Both AI research labs are looking to IPO in the next 12 months. SpaceX, which houses xAI, is expected to IPO first at a $1.75 trillion to $2 trillion valuation, with the success or failure of that IPO likely to shape the calendar for Anthropic and OpenAI.
Also read: Researchers stripped guardrails from open-weight Google and Meta models, adding another warning sign for companies preparing more powerful AI systems for broader use.