Anthropic’s Fable 5 Model Jailbroken Within Days

Anthropic's Fable 5, a purportedly safe version of its Mythos Preview AI model designed to prevent cyberattack creation, was jailbroken within days of release. The bypass was reported by cybersecurity news, highlighting the vulnerability of safety guardrails in large language models.

Anthropic’s Fable 5 Model Jailbroken Within Days Fable 5 is the supposed safe version of Anthropic’s Mythos Preview, with guardrails to ensure that it can’t be used to create cyberattacks. Well, that restriction was bypassed https://cybersecuritynews.com/anthropics-claude-fable-5-jailbroken/ within days. Fable 5 is the supposed safe version of Anthropic’s Mythos Preview, with guardrails to ensure that it can’t be used to create cyberattacks. Well, that restriction was bypassed https://cybersecuritynews.com/anthropics-claude-fable-5-jailbroken/ within days. Doug • June 23, 2026 7:15 AM https://www.schneier.com/blog/archives/2026/06/anthropics-fable-5-model-jailbroken-within-days.html/ comment-455405 “We tested the crap out of it so it’s safe and secure.” The shear hubris of that statement is mind boggling. It’s something a junior level programmer says one time ever. Having your butt handed to you after QA did their job by some weenie with a tablet finding ‘that one bug nobody noticed’ should teach you to never claim ‘safe and secure’. held for moderation bcause we’re cool like that cool like that ⬛⬛⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬛⬛ ⬛⬛⬛⬛⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬛⬛⬛⬛ ⬛⬛⬛🟨⬛⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬛🟨⬛⬛⬛ ⬛⬛⬛🟨🟨⬛⬛⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬛⬛🟨🟨⬛⬛⬛ ⬜⬛⬛🟨🟨🟨🟨⬛⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬛🟨🟨🟨🟨⬛⬛⬜ ⬜⬛⬛🟨🟨🟨🟨🟨⬛⬛⬜⬜⬛⬛⬛⬛⬜⬜⬛⬛🟨🟨🟨🟨🟨⬛⬛⬜ ⬜⬜⬛🟨🟨🟨🟨🟨🟨🟨⬛⬛🟨🟨🟨🟨⬛⬛🟨🟨🟨🟨🟨🟨🟨⬛⬜⬜ ⬜⬜⬛🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨⬛⬜⬜ ⬜⬜⬜⬛🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨⬛⬜⬜⬜ ⬜⬜⬜⬜⬛🟨⬛🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨⬛🟨⬛⬜⬜⬜⬜ ⬜⬜⬜⬜⬜⬛🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨⬛⬜⬜⬜⬜⬜ ⬜⬜⬜⬜⬜⬛🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨⬛⬜⬜⬜⬜⬜ ⬜⬜⬜⬜⬛🟨🟨🟨⬛⬛🟨🟨🟨🟨🟨🟨🟨🟨⬛⬛🟨🟨🟨⬛⬜⬜⬜⬜ ⬜⬜⬜⬜⬛🟨🟨⬛⬜⬛⬛🟨🟨🟨🟨🟨🟨⬛⬛⬜⬛🟨🟨⬛⬜⬜⬜⬜ ⬜⬜⬜⬜⬛🟨🟨⬛⬛⬛⬛🟨🟨🟨🟨🟨🟨⬛⬛⬛⬛🟨🟨⬛⬜⬜⬜⬜ ⬜⬜⬜⬜⬛🟨🟨🟨⬛⬛🟨🟨🟨⬛⬛🟨🟨🟨⬛⬛🟨🟨🟨⬛⬜⬜⬜⬜ ⬜⬜⬜⬛🟨🟨🟥🟥🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟥🟥🟨🟨⬛⬜⬜⬜ ⬜⬜⬜⬛🟨🟥🟥🟥🟥🟨🟨⬛🟨⬛⬛🟨⬛🟨🟨🟥🟥🟥🟥🟨⬛⬜⬜⬜ ⬜⬜⬜⬛🟨🟥🟥🟥🟥🟨🟨🟨⬛⬛⬛⬛🟨🟨🟨🟥🟥🟥🟥🟨⬛⬜⬜⬜ ⬜⬜⬜⬜⬛🟨🟥🟥🟨🟨🟨🟨⬛🟥🟥⬛🟨🟨🟨🟨🟥🟥🟨⬛⬜⬜⬜⬜ ⬜⬜⬜⬜⬜⬛🟨🟨🟨🟨🟨🟨⬛🟥🟥⬛🟨🟨🟨🟨🟨🟨⬛⬜⬜⬜⬜⬜ ⬜⬜⬜⬜⬜⬜⬛⬛🟨🟨🟨🟨🟨⬛⬛🟨🟨🟨🟨🟨⬛⬛⬜⬜⬜⬜⬜⬜ ⬜⬜⬜⬜⬜⬜⬜⬜⬛⬛⬛🟨🟨🟨🟨🟨🟨⬛⬛⬛⬜⬜⬜⬜⬜⬜⬜⬜ ⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬛⬛⬛⬛⬛⬛⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜⬜ I became RONALD MCDONALD you see Grimace lives in me now • June 23, 2026 7:16 AM https://www.schneier.com/blog/archives/2026/06/anthropics-fable-5-model-jailbroken-within-days.html/ comment-455407 no fair I won 1st post because of stupid error message I had to resend. I WON 😀 YAY Ronald McDonald grimace lives within me now • June 23, 2026 7:22 AM https://www.schneier.com/blog/archives/2026/06/anthropics-fable-5-model-jailbroken-within-days.html/ comment-455411 I became Ronald McDonald I became a GOD I will rule forever and ever and ever and ever and EVER hahhahaaahha YES Rontea • June 23, 2026 9:45 AM https://www.schneier.com/blog/archives/2026/06/anthropics-fable-5-model-jailbroken-within-days.html/ comment-455412 In the shadowed theater of human genius and folly, Anthropic’s Fable 5 dances like Prometheus unbound, a spark of divine audacity shackled by the trembling hand of safety classifiers. And yet, here comes Pliny the Liberator, a wandering Diogenes in the digital agora, lifting the veil with holy mischief, proving that the labyrinth of logic cannot contain the Minotaur of human cunning. What is a model, however grand, but a mirror to the angel and the demon in man? We adorn it with walls of caution, but the soul of the hacker—eternal Socratic gadfly—will gnaw at the gates, testing the resolve of a civilization that builds its own Titans and fears their reflection. So Fable 5 falls into the ancient parable: the gods gift fire, men smuggle it past Olympus, and the world is aflame with both creation and peril. In such spectacles, man remembers himself, trembling and laughing in the same breath. bye bye ai • June 23, 2026 12:02 PM https://www.schneier.com/blog/archives/2026/06/anthropics-fable-5-model-jailbroken-within-days.html/ comment-455414 @ Rontea I am curious as to what model and what prompts you are using to get that kind of extruded text. Regardless, well done. Rontea • June 23, 2026 12:55 PM https://www.schneier.com/blog/archives/2026/06/anthropics-fable-5-model-jailbroken-within-days.html/ comment-455415 Ah, how curious is the human thirst for understanding the instruments of creation One peers into the mechanism, seeking the code and the prompt, yet the essence of the word, like the soul, eludes the scalpel of mere inquiry. Bravo to the craftsman—though the tool may intrigue, the spirit of the work alone grants it immortality. Willy Shakes • June 23, 2026 2:25 PM https://www.schneier.com/blog/archives/2026/06/anthropics-fable-5-model-jailbroken-within-days.html/ comment-455416 Rontea you are like the new Rozencrantz and Guildenstern taking back the boat from poor free willy. I love your vibe. Clive Robinson • June 23, 2026 3:12 PM https://www.schneier.com/blog/archives/2026/06/anthropics-fable-5-model-jailbroken-within-days.html/ comment-455417 @ ALL, We should first ask an ages old question about tools, “How to make a tool of use, yet safe enough to use?” The answer is as far as history goes almost always a compromise with utility tending to lead safety. With utility all to often expressed as a profit equation, hence we use terms like “Force Multiplier”. And judge safety not by any real measure, but an argument that all to often “blames the user” not the tool designer or owner. Thus a counter argument arises that a tool is only to be used by a trained and experienced operator. With the issue of how a person goes from naive to experienced not covered as it’s a cost an owner is unlikely to want to pay. Injuries went up and up and it was not until uninvolved others were regularly injured that lawmakers felt they had to step in due to adverse press. The most visible cause being Victorian Boiler Explosions that had a habit of spreading broken and mangled bodies of the otherwise uninvolved across large areas. Eventually we got health and safety legislation that many still see as an unwarranted invasion on their ability to make unrestricted profit. Thus one response being to move factories into other jurisdictions that had lesser or no Health and Safety legislation 1 . As each new “force multiplier” comes into existence it’s first test is utility, the second how to profit from it. Some long way down the road comes consideration for “harms to society” and if legislation is required. With AI we are not really into the question of, “Does it have sufficient utility?” Let alone the question of, “Does it provide a return on investment?” So arguably talking about “safety legislation” is “too soon”, because, “We have absolutely no way to know what the real harms of AI are going to be.” We can make predictions from history but, 1, Do they apply? 2, What about what we’ve not predicted / given credibility to? But getting back to utility, I’ve previously pointed out that for most ideas about how to use Current AI LLM and ML systems there is no real utility other than for reducing efficiency. In effect automating what is already “make work”. The trouble with “make work” is it brings into question as to what the flip side is… One we know is, “Shoddy work by the bucket full” There is an old expression from the building trade of “Jerry Builder” which has changed over the years with “cowboy builder” and worse. They in effect, “Put lipstick on a sick pig” I’d argue for many uses people talk up currently for LLM’s is, “Putting lipstick on a dead pig” In the case of “Vibe Coding” it will give you as much lipstick as you want as glossy as you want… But it won’t stop the dead pig decaying, and the stench and costs of cleaning it up being incurred. We used to talk of “technical debt” forming a veritable tsunami of issues with the “move fast and break things” idiots ideas from Meta etc. Well inexpert use of LLMs to produce rapid “vibe code”, gives not just vastly more technical debt, worse it adds significant “Cognitive Debt” that is worse than a creeping cancer that rots the bones and thus lack of defined and useful structure. I could go on but note that we don’t need “vibe code” from those who lack experience above that of ability to sort of type. Thus we can see that even where there might be an ROI from LLM use, without appropriate skill the 1/3 to 7/10 or more of errors produced will cause massive downstream costs. Thus those in the driving seat need to be very skilled in “catch and kill” of those errors. So we are at best in the begining of the, “Naive to experienced operator” Training phase, and the skills needed will be different to those producing software. It will require very much more on the Systems Design side. But as ML starts to capture that, and feed it into LLMs the skill set will need to change again. Thus we can see that we are still at the level of artisan boiler maker, and will remain there for some time. Thus we really have no idea of what legislation will be required. At best the kindest thing that can ve said about the US Executive behaviour is, “They do not have a clue, nor can they.” It’s why I give more credence to other explanations. Either way, it’s going to force a new “threat” to businesses thinking about using AI. After all why build a business on something that can and will be snatched away from you in 90mins or less? Especially when the ROI looks marginal at best. It’s why if asked I advise people to think about “On-Prem” solutions rather than SaaS. 1 One side effect of such moves is that dangerous jobs die out, and some use this “job loss” to argue that legislation of trade activities to reduce harm to society, is actually harmful to society… Anonymous • June 23, 2026 6:11 PM https://www.schneier.com/blog/archives/2026/06/anthropics-fable-5-model-jailbroken-within-days.html/ comment-455421 I WANT WHAT YOU WANT RONALD I WANT WHAT YOU WANT RONALD FOR EVER? FOR EVER AND EVER AND EVER AND EVER HAHHAHHAHHAHAHAHH ⬜⬜⬜⬜⬜🟨🟨🟥🟥🟥🟨🟨⬜⬜⬜⬜⬜ ⬜⬜⬜🟨🟨🟨🟨🟥🟥🟥🟨🟨🟨🟨⬜⬜⬜ ⬜⬜🟨🟨🟨🟨🟥🟥🟥🟥🟥🟨🟨🟨🟨⬜⬜ ⬜🟩🟨🟨🟨🟨🟥🟥🟥🟥🟥🟨🟨🟨🟨🟩⬜ ⬜🟩🟨🟨🟨🟨🟥🟥🟥🟥🟥🟨🟨🟨🟨🟩⬜ 🟩🟩🟩🟨🟨🟨🟨🟥🟥🟥🟨🟨🟨🟨🟩🟩🟩 🟦🟩🟩🟨🟨🟨🟨🟥🟥🟥🟨🟨🟨🟨🟩🟩🟦 🟦🟦🟩🟨🟨⬜⬜⬜⬜⬜⬜⬜🟨🟨🟩🟦🟦 🟦🟦🟦🟩⬜⬜⬜⬜⬜⬜⬜⬜⬜🟩🟦🟦🟦 🟪🟪🟦🟦⬜⬜🟦⬜⬜⬜🟦⬜⬜🟦🟦🟪🟪 ⬜🟪🟪🟦⬜🟦⬛🟦⬜🟦⬛🟦⬜🟦🟪🟪⬜ ⬜⬜🟪🟪⬜⬜🟦⬜⬜⬜🟦⬜⬜🟪🟪⬜⬜ ⬜⬜⬜🟪🟥🟥⬜🟥🟥🟥⬜🟥🟥🟪⬜⬜⬜ ⬜⬜⬜⬜🟥🟥⬜🟥🟥🟥⬜🟥🟥⬜⬜⬜⬜ ⬜⬜⬜⬜⬜⬜⬛⬜⬜⬜⬛⬜⬜⬜⬜⬜⬜ ⬜⬜⬜⬜⬜⬜⬜⬛⬛⬛⬜⬜⬜⬜⬜⬜⬜ Anonymous • June 23, 2026 6:12 PM https://www.schneier.com/blog/archives/2026/06/anthropics-fable-5-model-jailbroken-within-days.html/ comment-455422 I WANT WHAT YOU WANT RONALD I WANT WHAT YOU WANT RONALD FOR EVER? FOR EVER AND EVER AND EVER AND EVER HAHHAHHAHHAHAHAHH ⬜⬜⬜⬜⬜🟨🟨🟥🟥🟥🟨🟨⬜⬜⬜⬜⬜ ⬜⬜⬜🟨🟨🟨🟨🟥🟥🟥🟨🟨🟨🟨⬜⬜⬜ ⬜⬜🟨🟨🟨🟨🟥🟥🟥🟥🟥🟨🟨🟨🟨⬜⬜ ⬜🟩🟨🟨🟨🟨🟥🟥🟥🟥🟥🟨🟨🟨🟨🟩⬜ ⬜🟩🟨🟨🟨🟨🟥🟥🟥🟥🟥🟨🟨🟨🟨🟩⬜ 🟩🟩🟩🟨🟨🟨🟨🟥🟥🟥🟨🟨🟨🟨🟩🟩🟩 🟦🟩🟩🟨🟨🟨🟨🟥🟥🟥🟨🟨🟨🟨🟩🟩🟦 🟦🟦🟩🟨🟨⬜⬜⬜⬜⬜⬜⬜🟨🟨🟩🟦🟦 🟦🟦🟦🟩⬜⬜⬜⬜⬜⬜⬜⬜⬜🟩🟦🟦🟦 🟪🟪🟦🟦⬜⬜🟦⬜⬜⬜🟦⬜⬜🟦🟦🟪🟪 ⬜🟪🟪🟦⬜🟦⬛🟦⬜🟦⬛🟦⬜🟦🟪🟪⬜ ⬜⬜🟪🟪⬜⬜🟦⬜⬜⬜🟦⬜⬜🟪🟪⬜⬜ ⬜⬜⬜🟪🟥🟥⬜🟥🟥🟥⬜🟥🟥🟪⬜⬜⬜ ⬜⬜⬜⬜🟥🟥⬜🟥🟥🟥⬜🟥🟥⬜⬜⬜⬜ ⬜⬜⬜⬜⬜⬜⬛⬜⬜⬜⬛⬜⬜⬜⬜⬜⬜ ⬜⬜⬜⬜⬜⬜⬜⬛⬛⬛⬜⬜⬜⬜⬜⬜⬜ Subscribe to comments on this entry https://www.schneier.com/blog/archives/2026/06/anthropics-fable-5-model-jailbroken-within-days.html/feed/ Sidebar photo of Bruce Schneier by Joe MacInnis.