Artificial Intelligence & Machine Learning , Critical Infrastructure Security , Governance & Risk Management

The abrupt, government-ordered cut-off of access to Mythos 5, the most cyber-capable of Anthropic's large language models, has underlined a message security experts have been trying to get out to the operational technology community: You don't need Mythos.

See Also: Edge Transformation: Top 5 SASE Predictions and Trends

"If there's any one message I've been trying to get out, it is, 'No, you don't have to wait to do this,'" said Patrick Miller, CEO of Ampyx Cyber, an OT security consulting firm.

Any of the commercially available large language models can be used to find vulnerabilities in software, Miller told ISMG in an interview. Mythos, and especially the latest version, Mythos 5, might be better at it than the others, but "pretty much any of these LLMs can do that fairly well," he said.

The underlying model is only one factor in determining how effective such vulnerability hunts will be, Miller added. The software infrastructure that wraps around the LLM, known as a harness, and special training on the codebase they were investigating and other contextual matters, are also important.

"Train them on your code, train them on your development life cycle and just go to town," he said. "You don't need to wait to be part of Glasswing," Anthropic's invitation-only club, which allows a relatively small number of companies access to Mythos, which the frontier lab deems too dangerous for public release (see: Anthropic Limits on OT Access to Mythos Draw Criticism).

The Trump administration on Friday used export control authorities to direct Anthropic to limit access to Mythos 5, and its publicly available version, known as Fable 5, to U.S. citizens only. That meant that foreign Anthropic employees working in the United States would have been prohibited from using the models, and the company would have to find a way to check the citizenship of any potential user. Anthropic decided to remove the models from service completely.

The move reportedly followed the successful jailbreak of Fable 5 by a researcher from Amazon. Fable 5 was released with guardrails - software designed to keep it from being used offensively by hackers, for instance by routing certain kinds of queries back to earlier, less capable models. Finding ways to get the model to work around those restrictions is known as jailbreaking.

Anthropic did not respond to a request for comment and it wasn't immediately clear what impact the move had on Glasswing, although the company said none of their other models were affected, so it seemed likely Glasswing members would simply be rolled back to using Mythos 4.

Miller's sentiments that companies shouldn't wait to start using LLMs to look for vulnerabilities in their systems were widely echoed in the OT security community.

The vast majority of companies making software, especially in the OT sector, did not have access to Mythos. "Waiting for this model, or any other, is not the way to go," said Moreno Carullo, CTO of Nozomi Networks, a pure-play OT security firm.

"The way to go is to make sure you are stress testing your software against a current model," he added, "I can tell you it will find a lot of [vulnerabilities], and there will be a lot of work to do."

The idea behind Mythos was to allow companies a head start, to use the power of the new model to find software vulnerabilities and give companies the chance to fix them before more capable models become widely available - including to hackers.

"Companies which haven't stress tested [their codebase] are already behind," Carullo told ISMG.

Research shows that artificial intelligence vulnerability discovery capability is "jagged," varying widely depending not just on model size and sophistication but a host of other factors. Aisle, a company specializing in using AI for vulnerability discovery, tested Mythos against other, smaller, open-source models.

"The capability rankings reshuffled completely across tasks. There is no stable best model across cybersecurity tasks. The capability frontier is jagged," the company wrote.

Eight "small, cheap, open-weights models" were all able to find the "flagship" Mythos-discovered vulnerability in OpenBSD, Aisle found.

Other experts suggested that LLMs could help security teams at OT and other companies in many other ways, not just with vulnerability discovery.

"Where we're getting it a little bit wrong is we're hyper-focusing on Mythos and … just on vulnerability discovery," said Victor Wieczorek, senior vice president of offensive security at GuidePoint Security. "We're looking at them through a very myopic vulnerability lens, and that's not the full story."

He said his own team were using LLMs to "remove all sorts of drudgery around the work that they do. … Writing hundreds of page reports and then translating them to technical findings and then analyzing and reviewing" them.

OT teams could use LLMs "to remove other drudgery not related to vulnerability, not related to downtime, not related to patching, right? Other things in their work days that they can remove using these tools."

But the asymmetry in the way LLMs work - making them better at finding vulnerabilities than at fixing them - troubled some experts.

"We've spent the last 40 years developing this very collaborative, integrated way of sharing information about vulnerabilities and now new barriers to entry are being erected," said Matthew Butkovic, technical director for cyber risk and resilience at the Software Engineering Institute at Carnegie Mellon University.

"These new frontier models are going to enumerate vulnerabilities at a pace we've never seen before," he said. "They're really good at finding issues. They don't offer a comparable level of ability when it comes to resolution. They're not going to fix those things with the same speed. That's another challenge here."

"The truth is," Butkovich concluded, "everyone is rushing at full speed for market share, and to get new and novel [capabilities] to market, but I think that we've got to balance that with the remediation required once you build something that might have a high potential for misuse and abuse, and could be potentially dangerous."

Senior technical Anthropic staff flew to Washington at the weekend, reported Axios, to meet with officials from the White House, CIA and Commerce Department to try to repair the dispute.