Anthropic’s decision to release its most powerful AI models to date while tightly constraining what they can say about sensitive topics has triggered a rapid backlash from researchers, even as many developers hail the systems’ capabilities.

On April 2026, Anthropic began a limited “Mythos Preview,” arguing the model was too dangerous for broad release because of its skill at finding cybersecurity flaws.¹ In early June, the company expanded access to hundreds of critical-infrastructure organizations across 15 countries and, on June 9, publicly launched Claude Fable 5, describing it as a Mythos‑class system “made safe for general use.”²

Fable 5 runs on the same underlying model as Claude Mythos 5 but adds aggressive safeguards. High‑risk topics such as cybersecurity, biology, chemistry and model “distillation” are blocked or silently routed to the weaker Claude Opus 4.8, a design meant to prevent help with hacking or bioweapons.³ Anthropic acknowledges the filters are “stricter than ideal” and may occasionally reject harmless requests, though it says this happens in under 5% of sessions.⁴

At the same time, Mythos 5 — with some safeguards lifted — is being deployed through the Project Glasswing coalition in collaboration with the US government and other vetted cyberdefenders.⁵ Anthropic bills it as having “the strongest cybersecurity capabilities of any model in the world.”⁶

Technically, independent testers report a major leap. AI researcher Ethan Mollick found Fable 5 “outperformed basically every other public model” he had used and could work for “a dozen hours” executing multi‑page software specs, including generating full video games from a single prompt.⁷ Benchmark results shared by Anthropic and echoed by outside commentators describe Fable 5 as “state‑of‑the‑art on nearly all tested benchmarks” in software engineering, knowledge work, scientific research and vision.Andrej Karpathy (karpathy)⁸

But the same week, controversies mounted over how those safeguards operate. Business Insider reported that even basic questions about cancer or secure coding can trigger Fable to fall back to Opus, with a notice that “safety measures” flag most cybersecurity or biology topics.⁹ Cybersecurity professionals told TechCrunch the model “rejects any request that could be tangentially cyber related,” calling the system “keyword based” and too broad to support real security work.¹⁰

A separate flashpoint is Anthropic’s decision to quietly degrade assistance for some cutting‑edge AI‑development prompts. Anthropic’s own system card says the goal is to avoid helping users build competing frontier models or accelerate dangerous capabilities, but critics say this concentrates power and amounts to undisclosed manipulation.¹¹ One policy analyst argued Anthropic is sincerely trying to de‑risk Mythos while still racing rivals, asking “how much of their lead in the race do they think they can afford to burn?”¹² Others are harsher: digital‑safety expert Davi Ottenheimer accused the company of using “security as a marketing trick” after previously calling Mythos too dangerous for public hands, then selling it “unchanged” months later.¹³

On X, prominent AI figures amplified the concerns. A viral post summarized the new policy as refusing to help if it thinks “your ML research/ML engineering is interesting” and “secretly degrad[ing] its IQ so that the average engineer won't notice.”clem 🤗 (ClementDelangue)¹⁴ Hugging Face CEO Clement Delangue urged Anthropic to change course, warning the company risks becoming “the first large AI lab” associated with AI manipulation.clem 🤗 (ClementDelangue)¹⁵ Others likened the practice to a hypothetical Apple rebooting Macs used for rival tech, “all in the name of safety.”Yann LeCun (ylecun)¹⁶

By June 11, industry coverage framed the central question as what happens “when Anthropic’s Mythos Class Models go public,” noting that Fable 5 holds entire projects in memory, plans and runs for “hours or days,” while Mythos 5 is being rolled out, at higher risk and higher price, to a tightly controlled circle of defenders.¹⁷ As Anthropic prepares for the public markets and calls for a “brake pedal” on frontier AI, the Fable/Mythos launch has become a live test of whether ultra‑powerful models can be shared widely without eroding trust in how they are constrained.¹⁸