Generated by Codex with GPT-5

What this article is about

This summary covers The Economist’s April 11th, 2026 Business article listed in the contents as A dangerous new AI and published under the headline Mythical monster.

The article argues that warnings about AI danger are usually worth treating with skepticism, because frontier labs have incentives to dramatise their own progress. This time, though, The Economist thinks the warning may be genuine. Its focus is Anthropic’s latest model, Mythos, which the company says is powerful enough at discovering and exploiting software flaws that it should not yet be released broadly.

Why this warning lands differently

The piece begins by recalling an earlier episode from 2019, when OpenAI briefly held back GPT-2 on the grounds that it was too dangerous to release. That caution ended up looking overblown. Much stronger models followed without producing the kind of catastrophe that had been implied. The article uses that history to show why people should normally be careful about taking AI-company alarm bells at face value.

Anthropic’s claim, however, is presented as more specific and more credible. Rather than gesturing at vague social harms, the company is pointing to a narrow and technically serious capability: the ability to locate major vulnerabilities in widely used software, then either fix them or weaponise them. The Economist suggests that this kind of capability is much easier to imagine causing real damage than the more abstract fears that often accompany new model launches.

The case for taking Anthropic seriously

The article offers two main reasons for believing Anthropic may not be exaggerating. The first is the severity of what it says the model has already done. According to the company, Mythos has identified serious flaws across major operating systems and web browsers, including one that had reportedly remained unnoticed for decades. If true, that moves the issue from speculative risk into something much more immediate.

The second reason is the behaviour of other firms. Anthropic is not acting alone. It has paired the pause in wider release with Project Glasswing, an effort to let major software companies use the model to harden their systems before the technology spreads more broadly. The article notes that participants include important industry players, even some that compete directly with Anthropic. That outside participation makes the threat look less like marketing theatre and more like a shared industry judgment that the model’s offensive cyber capabilities are unusually strong.

Security first, business too

The article is careful to note that a security-minded release strategy also serves Anthropic’s commercial interests. The company will initially subsidise use of the model in Project Glasswing, but plans eventually to charge substantially more for it than for its previous flagship. In other words, caution and product strategy are not separable here. Anthropic gets to present itself as the responsible actor while also building demand for a premium system.

That does not mean the safety argument is insincere. The Economist’s point is subtler: even when a company is behaving prudently, it is still making high-stakes choices inside a market structure full of incentives, rivalries and branding opportunities. The governance problem is not solved simply because one lab sounds more careful than the others.

The strategic complication

The most interesting part of the article is that the risk is not confined to criminals. If a model becomes exceptionally good at surfacing hidden flaws, it may also disrupt the way governments use cyber-offence. America has long kept some undisclosed software vulnerabilities in reserve so they can be exploited later for intelligence or military purposes. A system that helps software companies find and patch those weaknesses early could blunt part of that arsenal.

That creates a conflict between two legitimate goals: making the digital world safer overall, and preserving offensive tools for state use. The article hints that this tension helps explain why Anthropic has clashed with parts of the American national-security establishment before, and why its attempt to close vulnerabilities faster may have political as well as technical consequences.

The takeaway

The article’s broader claim is that AI risk is becoming more concrete. Instead of arguing about distant superintelligence or generic disruption, companies and governments may now have to deal with a model that can materially shift the balance between attackers and defenders in cybersecurity. The immediate question is not whether such systems are powerful in the abstract. It is who gets access first, how long defenders have to prepare, and whether that window is enough.

In plain English: The Economist sees Anthropic’s new model as one of the first cases where “too dangerous to release” may be more than hype. The deeper lesson is that frontier AI is starting to create practical security dilemmas that blend technology, business incentives and state power into the same decision.