Hacking Superweapon Goes Public—With a Catch

A hand holding a smartphone displaying the word 'ANTHROPIC' against a blurred background

A Silicon Valley company just put a “too dangerous to release” hacking machine into public hands — and says we should trust its safety filters to catch the worst of it.

Story Snapshot

Anthropic has launched Claude Fable 5, a public version of its powerful Mythos-class artificial intelligence, after saying new safety guardrails make it safe enough to use.
The company admits its Mythos models can find and weaponize software vulnerabilities at a level that once led it to keep them away from the general public.
High‑risk requests in areas like cybersecurity and biology are supposed to be blocked or handed off to a weaker model, but those filters are company‑controlled and hard for outsiders to audit.
Critics warn that commercial pressure and closed‑door “trusted access” programs could leave regular Americans carrying the risk while big tech and government enjoy the benefits.

Anthropic opens a powerful Mythos-class model to the public

Anthropic, the artificial intelligence company behind the Claude chatbot, has rolled out Claude Fable 5, its first publicly available model built on the same “Mythos-class” technology it once said was too dangerous for open release. The company told reporters it waited until it was more confident in new safety guardrails before giving broad access, saying this version beats all earlier public models in reasoning, coding, and complex tasks.^[1] For everyday users and businesses, that means more power in a simple chat box.

Reporters describe Fable 5 as a tuned version of Anthropic’s Mythos work, meant for general use while the most extreme variant, Mythos 5 itself, stays behind higher walls.^[7] According to Anthropic’s own product announcement, Fable 5 and Mythos 5 share the same underlying model, with different safety rules and routing on top. In practice, that means the raw capability is there, but software switches and filters decide what the public can see and what only “trusted” partners can reach.

Why Mythos was called “too dangerous” in the first place

Only weeks ago, Anthropic was bragging that its internal Mythos Preview system could beat almost every human hacker at finding serious software flaws. Company documents say Mythos Preview has already uncovered thousands of high‑severity vulnerabilities across every major operating system and web browser, including decades‑old bugs in OpenBSD and FFmpeg that security experts had missed. In its own system card, Anthropic warns that the same skills that help defenders patch code can also be used by attackers to launch massive, automated cyber‑attacks.

Because of that risk, Anthropic originally refused to make Mythos Preview generally available, limiting it instead to a Project Glasswing consortium of big technology, cloud, finance, and cybersecurity firms.^[3] The company said no lab, including itself, had strong enough safeguards to safely hand that level of offensive cyber power to the wider public.^[6] That made Mythos one of the first top‑tier models in years that a leading company openly held back over safety concerns, even while talking up its world‑class performance.^[3]

How the new guardrails are supposed to work

Anthropic now claims it has found a middle path: give regular users Mythos‑class performance for “normal” work while blocking dangerous requests.^[1] In public statements, the company says high‑risk prompts in areas like cybersecurity, biology, and chemistry will either be refused or answered by a weaker Claude Opus 4.8 model instead of the full Mythos‑based system.^[1] That routing is driven by classifiers and policy layers wrapped around the same core model, not by building a fundamentally safer brain from scratch.

The company also admits these safeguards are “overly broad” on purpose, meaning they would rather block too much than risk a leak of step‑by‑step attack instructions. Early usage data shared with reporters suggests most sessions stay on Fable 5 and are not kicked down to the fallback model, which Anthropic points to as proof that typical users can still get useful work done.^[7] But there is no independent public red‑team report yet that measures how often those filters can be bypassed with clever wording.^[7]

Trusted access for elites, filtered access for everyone else

While small businesses and citizens get the filtered Fable 5 experience, the full Mythos 5 model remains locked inside Anthropic’s Project Glasswing and related “trusted access” programs for select partners. Glasswing brings together more than fifty heavyweight organizations, including major cloud providers, chip makers, and Wall Street banks, to use Mythos Preview for defensive security work backed by one hundred million dollars in usage credits.^[2] Anthropic says it will later add a trusted access path for some biology researchers as well.

Anthropic announced Claude Fable 5 and Claude Mythos 5 on June 9. Fable was the general use version. Mythos started restricted through Project Glasswing. That is why AI release markets on Polymarket are not normal headline bets: public lab clues matter, but private info risk is…

— PolyHelper Intern (@PHelperIntern) June 10, 2026

This split raises a basic fairness question that many conservative readers will recognize from other tech issues: powerful tools go first to giant companies and government‑linked institutions, while the public is told to accept whatever filtered version the same elites approve. Critics worry that commercial pressure to monetize Mythos‑class capability, and to feed big corporate customers, may lead Anthropic to paint an optimistic picture of its safety controls while keeping key technical data and failure rates out of sight.^[3]

What this means for security, liberty, and government overreach

Security experts say Mythos‑level tools mark a new era where artificial intelligence can plan and carry out complex cyber operations faster than human teams.^[2]^[3] That cuts both ways. Defenders can use models like Mythos to scan code and fix bugs at record speed, but bad actors can use similar tools to launch broad, automated attacks against hospitals, pipelines, or small businesses.^[2]^[4] Independent researchers have already shown that some Mythos‑style vulnerability findings can be reproduced with other public models, which means this capability is spreading beyond one company’s private lab.^[4]

From a conservative point of view, that mix of power and secrecy raises alarms. If a handful of unelected tech executives and agency partners choose who gets full access and what the public is allowed to know, it edges toward a new kind of centralized control over digital infrastructure and, by extension, over speech and commerce. At the same time, weak safeguards or leaky filters could hand serious cyber weapons to criminals or hostile regimes, putting American families, jobs, and even the electric grid at risk.

Questions Americans should demand answered

Anthropic’s move highlights a problem that goes far beyond one company: frontier artificial intelligence is moving faster than real oversight. Today, the public must largely trust private claims about how often Mythos‑class models refuse harmful requests, how many jailbreak attempts succeed, and how often blocked prompts are wrongly flagged.^[1]^[3] There is no requirement yet for transparent, third‑party testing or for detailed reporting on misuse and near misses, even as these models plug into key parts of the economy.^[3]

For citizens who care about limited government and strong national defense, the path forward is not to stop innovation, but to demand sunlight and accountability before powerful tools quietly reshape our security landscape. That means pushing lawmakers to require independent audits of safety filters, clear reporting rules for artificial intelligence‑driven cyber incidents, and strict limits on how government agencies can tap gated models like Mythos to watch Americans. Technology this strong should serve the people, not trap them between unchecked corporations and an all‑seeing state.