Mythos, the NSA and the game of 'telephone': when an AI capability is amplified until it's unrecognizable

Few stories better illustrate the fragility of our information ecosystem around artificial intelligence than that of Anthropic's Mythos model and its tests with the NSA's systems. Webb Wright's article in Gizmodo, published on June 24, 2026, does not add fuel to the fire: it does the opposite, reconstructing step by step how a claim was distorted until it became something qualitatively different from what actually happened. And that clearing work is today as valuable as the technology news itself.

The account begins in April 2026, when Anthropic unveiled Mythos and described it as extraordinarily effective at identifying and exploiting software vulnerabilities. Its potential seemed so delicate that the company chose not to release it generally and to limit access to a small group of evaluators, among them the NSA. Weeks later it emerged that the agency had detected multiple flaws in its own systems during those tests, and the question leapt like a spark: if the best-protected organization in the world is vulnerable, what awaits the rest?

This is where it is worth pausing. According to the article, the panic originated in a report by The Economist that picked up statements by Senator Mark Warner at a hearing on June 11; Warner claimed that Mythos had penetrated «almost all the classified systems» of the NSA «not in weeks, but in hours», attributing the figure to the agency's own director, General Joshua Rudd. The combination of a top-level source, a dramatic framing and the outlet's reputation created the perfect breeding ground for uncontrolled spread. Gizmodo aptly describes it as a game of broken telephone.

The context lost along the way is precisely what dismantles the sensationalism. A later report by the New York Times, cited in the article, specified that the tests were carried out in a digital environment so controlled that it would be nearly impossible for an external attacker to replicate. And, above all, that Mythos identified vulnerabilities but, according to those officials, at no point came to exploit them. In cybersecurity that line is everything: detecting a flaw is diagnosis; actively taking advantage of it is the attack. Confusing the two turns a potentially useful defensive tool into an imaginary apocalyptic weapon.

The piece also includes an unusual gesture: the author of the original text in The Economist himself acknowledged on X that his portrayal had been misleading, admitting that he quoted Warner «to give an idea of Mythos's power» and that «it was a mistake not to have added nuance». That public self-correction, far from being an anecdote, should be read as a professional reminder that data without context does not inform: it deforms.

A dense business and political context orbits the episode. According to the article, Anthropic has reportedly displaced OpenAI as the world's most valuable startup and is said to be preparing for a historic IPO, so that Mythos's aura also functions as a reputational asset. In parallel, the text notes that the Trump administration ordered in early June that foreign nationals' access to Fable 5 —a model of the same class— be restricted, invoking export-control legislation, an interpretation that several legal scholars reportedly described as «spurious» and that cybersecurity experts considered counterproductive for U.S. defenses themselves. These assessments should be attributed to their sources and not taken as proven.

The moral is calm and applicable to almost any headline about AI capabilities: before getting scared, it is worth asking who said it, under what conditions it was tested and whether detecting has been confused with exploiting. Technology advances quickly; the quality of our interpretation does not always keep pace. And in that gap, more than in the models, is where avoidable panics are incubated.