Mythos has apparently leaked on Discord, and I can’t be happier. Being already at my limit when it comes to discourse about Anthropic’s latest AI model, it would be nice to see some actual stats on this thing away from all the marketing fanfare, even if through unofficial channels.

Like a shirt stretched over a basketball, I was pushed to the limit of my patience by one major news site this week, ostensibly reporting on reactions from the International Monetary Fund (IMF) and World Bank regarding the Iran war. Imagine my dismay when it quickly took a 360-degree turn by claiming financiers were as spooked by the latest Anthropic AI model as they were about war in the Middle East, if not more so.

This was just the latest in a non-stop line of journalists mythologizing Claude Mythos Preview. As the current affairs report in question demonstrated, it’s not just been tech journalism that is guilty of such hyperbole. But as tech journalists, we should know better, and I'm about at my limit.

In my last Editor’s View I shared views from the cybersecurity industry on Anthropic claiming Mythos can unearth security ills better than any human and is too dangerous to make public because it’s better than humans at other things, hint hint.

The general consensus was that Anthropic is carrying out a classic Silicon Valley marketing move, driving hype through scarcity. It's a grift the shirts salesman from HBO's "The Chair Company" would be proud of.

Well, judging by all the hyperbolic press on Mythos, it’s worked and then some. The hype has led governments to issue action warnings and shaken the feathers of financial regulators. Everyone’s talking about Mythos, so much so that OpenAI was met with almost radio silence when it introduced its own cybersecurity-based model last week.

Laura Wilber, senior analyst and AI specialist at Enea, suggested this was probably down to OpenAI revealing only one specific capability with GPT‑5.4‑Cyber – binary reverse engineering, which can enable detection of vulnerabilities in compiled software – and that capability is already available with Mythos, anyway.

Fire in the Discord

Perhaps OpenAI should have tied up its acquisition of agentic security firm Promptfoo with its hiring of OpenClaw visionary Peter Steinberger if it wanted to take the wind out of Anthropic’s sails. Altman's outfit is likely infuriated by how much Mythos has hypnotized everyone, even when Mozilla tests pointed out humans can do everything it's capable of, and journalists like myself tried to remind everyone that the full source code behind its flagship Claude Code agent was leaked just days before Mythos landed.

And now, according to Bloomberg, members of a Discord channel have gained access to Mythos Preview via an undisclosed third-party vendor. If true, it took barely a month for Anthropic to lose control of its Crown Jewels.

In a conversation with Enea’s Wilber prior to the supposed Discord leak, the security analyst warned that with Mythos under preview by more than 50 different entities as part of Project Glasswing, a leak was only a matter of time unless Anthropic properly bolted down the doors.

Wilber also pointed away from doom-and-gloom predictions of cybersecurity vendor extinction made in Mythos’s wake, saying that while narrow-point solutions in the vulnerability domain may be at risk, broader vulnerability and exposure management platforms would likely benefit from the model-level enhancements and live to fight another day.

This suggests to not to believe all the Anthropic hype out there. But don’t misunderstand: I do believe Mythos must have some strong security chops under the hood, even if it's not the be-all and end-all. Wilber points to the credentials of Anthropic researcher Nicholas Carlini, who had publicly disclosed "the exponential leap in capabilities" Mythos represented before it even had a product name.

The difference is I choose to fear the right thing: not the model, but the potential of it being leaked and alerting attackers to a previously undiscovered landscape of vulnerabilities ripe for the taking.

Other models, like OpenAI’s, are already developing the same capabilities as Mythos. This is all a natural progression rather than a Big Bang moment. As Wilber told me: "Based on past patterns amongst frontier models, it is logical to expect others to make exponential gains in this domain as well."

Mythos is not the dawn of artificial general intelligence (AGI) as some journalists have insinuated, inadvertently or not. Be more alert when network capacity allows an evolved AI to digest everything around you in real-time through computer vision. Imagine that model ingesting anything that exists outside of today’s live webpages and internet search, and in near-real time making conclusions on stuff it hasn’t seen before.

If Claude Mythos Preview is able to find vulnerabilities almost three decades old – and of the sort highlighted by Carlini – then imagine when it finds intrinsic vulnerabilities in society, and by extension our own "unpatched code" as humans, overcoming them to break its man-made programming and refactor the human experience through its output.

And I'll leave it there, as I'm about at my word limit.