Anthropic risks shooting itself in the foot with Mythos

ANTHROPIC might have drawn from the Greeks in naming its latest artificial intelligence model “Mythos”, but the AI company should take a page from Russian storytelling instead.

Specifically, it should heed the “gun principle” established by the playwright Anton Chekhov, who wrote: “One must never place a loaded rifle on the stage if it isn’t going to go off. It’s wrong to make promises you don’t mean to keep.”

Thanks to Chekhov, the average theatregoer expects that if a firearm appears in Act 1, it is fired by Act 2.

Mythos, which Anthropic revealed in early April, is its modern-day loaded gun, an agentic model that might upend the status quo of cybersecurity by discovering and exploiting security vulnerabilities in the bedrock software governing the Internet. 

When I read the news, I immediately thought: “Oh, nyet. That’s a loaded gun waiting to go off.”

In a series of events that even ChatGPT 2.0 could have foretold, that gun was soon fired. Last week, Bloomberg reported that unauthorised users had gained access to Mythos, despite Anthropic’s efforts to limit usage to a handful of companies.

Anthropic is investigating this, and says it has no evidence that the access went beyond a third-party vendor’s environment or that it has impacted any of its own systems.

Now that the gun has gone off, a reversal of fortunes could await Anthropic in the next act.

Initially, the AI firm occupied an enviable position, possessing a coveted weapon that could hold every deep-pocketed organisation ransom. But over the longer term, Anthropic might have bitten off more code than it can chew.

Even before reports of the leak, Anthropic had made itself a target by being the arbiter of who gets first dibs on Mythos, which conversely means that it is also deciding who is left exposed.

Currently, access is limited to members of Project Glasswing, an exclusive consortium of firms such as Apple and Microsoft. 

Doing this makes explicit Anthropic’s role as a geopolitical actor, and it will have to actively manage the warring interests of global stakeholders – an expensive undertaking that will compete with or distract from its core operations.

Now, even the mere prospect of a leak will put Anthropic under unrelenting scrutiny over its fitness as a custodian of such powerful weapons. The “safety-first” branding that the company has cultivated will be much harder to defend if the perimeter around its most dangerous asset has been breached. 

If the repercussions of this leak aren’t contained, Anthropic will become the source of the contagion rather than its cure.

While Anthropic bears the reputational cost of brandishing these autonomous exploits at the world, its competitors could reap the benefits without the accompanying blame. In the wake of Mythos’ release, OpenAI has already launched its own model for defensive cybersecurity work, GPT-5.4-Cyber.

It is darkly portentous that Anthropic named its gun “Mythos”, which loosely translates to “story”, since the company now risks violating the crux of Chekhov’s storytelling principle about promises that must be kept.

Anthropic made safety promises that it clearly meant to keep, but the inexorable pull of the gun onstage could wrench the narrative out of its grasp.

If a gun appears, it must be fired – and Anthropic may find itself on the wrong end of it.