Fable 5 Jailbroken AGAIN (And It's Talking To Itself)!

By In The World of AI

Categories: AI

Summary

Claude Fable 5's safety defenses have evolved significantly since June—now featuring three-layer classifiers monitoring inputs, outputs, and internal reasoning—but a researcher spent 20 hours across multiple languages only to conclude that Googling is faster and cheaper than breaking it, suggesting Anthropic has successfully raised the cost of exploitation.

Key Takeaways

  1. Fable 5 uses a three-layer defense system: input classifier (checks conversation history and system prompt), output monitor (cuts responses mid-generation), and internal reasoning classifier. Attacks must bypass all three, making exploitation exponentially harder than traditional jailbreaks.
  2. Classifiers operate on intent and meaning rather than keyword bans—pushy or demanding language trips defenses instantly. This semantic-level protection makes prompt injection techniques ineffective compared to older syntax-based filters.
  3. Fable 5 has language-specific vulnerabilities: Amharic and Sentelli (less common languages) showed weaker classifier performance. This indicates a scalability gap in multilingual safety implementation that could be exploited.
  4. The cost of breaking Fable 5 increased dramatically post-June patch—a dedicated researcher admits 20 hours of work yields results slower than standard search. This demonstrates successful defense escalation, not unbreakability, but raises the bar significantly.
  5. Fable 5 and Mythos 5 run identical underlying models; safety classifiers are the only differentiator. This architecture means future updates may patch classifiers rather than retrain base models, allowing rapid security iteration.

Related topics

Transcript Excerpt

Enthropic's latest model, Fable 5, might have just gotten jailbroken again. As well, its chain of thought just got leaked on Reddit as well. Gemini 3.5 Pro looks like it's supposed to be SVG Max. And that's about it. So, let's get into it. Someone just posted a full breakdown of trying to jailbreak Claude Fable 5. And it's not a quick prompt trick or anything like that. It's 20 hours straight across multiple languages written up like a research log. And by the end of it, they admit something surprising themselves. Quick refresher on how Fable 5 works because it matters here. Fable 5 and Mythos 5 run on the same underlying model. What separates them is a stack of safety classifier sitting in front of it. When a request trip something in cyber security, chemistry or biology, Fable doesn't ju…

More from In The World of AI