Fable
Anthropic
TLDR: âAnthropic debuted Claude Fable 5, calling it Mythos class and safe for general use. It routes sensitive queries to Opus 4.8, uses Project Glasswing testing, and prices it at $10 input tokens and $50 output tokens.
Key Takeaways:
- Anthropic paused Mythos Preview after it showed strong security vulnerability finding potential through trusted testing.
- Claude Fable 5 targets general release as a Mythos class model, with Project Glasswing safeguards and routing sensitive cybersecurity biology chemistry queries to Opus 4.8.
- Guardrails cut self handling to about 95 percent and followed 1,000 hours of bug bounty style testing without a universal jailbreak, with 30 day retention.
This is a classic AI tension: make the model scary useful, then bolt on enough brakes to keep it from becoming the next hacking shortcut. Anthropic is betting that smarter routing plus long jailbreak pressure will let the power ship safely.
This is a classic AI tension: make the model scary useful, then bolt on enough brakes to keep it from becoming the next hacking shortcut. Anthropic is betting that smarter routing plus long jailbreak pressure will let the power ship safely.
Q&A
If Fable 5 punts sensitive requests to Opus 4.8, could attackers still learn useful exploitation patterns indirectly through partial answers?
The safeguard design assumes the lower model provides accurate assistance without exploitation capability. Attackers may try to probe boundaries to elicit higher risk detail, which is why continuous testing and monitoring matter.
Why did Anthropic limit biology and chemistry at the model level instead of relying only on output filtering?
Anthropic points to dual use potential, including gene therapy enablement and distillation attempts. Blocking earlier reduces the chance that the model helps translate requests into actionable experimental pathways.
What does 1,000 hours without a universal jailbreak suggest about the rest of the attack surface?
It suggests no single bypass trick dominated during testing, but it does not guarantee zero jailbreaks. Attackers often pivot to new prompt chains, so ongoing red teaming and patch cycles remain essential.
How could longer tasks benefit users while still raising the risk of misuse for high capability models?
Anthropic says Fable 5 pulls ahead on longer and more complex work, which helps legitimate research and engineering. The same strength can also amplify harmful workflows, so routing and restricted domains become the control lever.
What changes if competitors copy the pricing and architecture approach but tune safety differently?
A model that is similar in capability can still differ sharply in exploitability if its safety stack and test coverage vary. The market may reward raw benchmarks first, then safety proof becomes the differentiator after incidents.
Top in AI
Anthropic opens Mythos model, but guards cybersecurity access
Hyattâs sales AI buys time, boosts group bookings
Europe wants sovereign AI, but chip control bites
iOS 27 pressures Apple Passwords to fix breaches automatically
No comments yet. Be the first to share your thoughts!