Criticism Mounts for Anthropic’s New Model ‘Fable’! Overly Strict Guardrails Make ‘Code Reviews Impossible’

News Summary

• Launch of the New Model “Fable”: Anthropic has unveiled “Fable” as a limited release of its powerful cybersecurity model “Mythos,” but experts are expressing numerous concerns.
• Excessive Guardrails: Reports indicate that the model overly reacts to keywords related to cybersecurity and biology, blocking even harmless activities like browsing blogs or conducting code reviews.
• Fallback Mechanism: When the guardrails are triggered, the model automatically switches to “Claude Opus 4.8.”

Key Points

• Keyword-Based Restrictions: Experts claim that even the mere inclusion of cybersecurity-related vocabulary can interrupt chats, effectively barring safe software engineering practices, such as secure coding.
• Distinction from Mythos: The top-tier model “Mythos” is provided only to select organizations through “Project Glasswing,” while the general public version “Fable” faces criticism for its seemingly arbitrary limitations.
• Verification Program in Place: Anthropic offers a “Cyber Verification Program,” allowing approved experts to access less restricted features, yet the inconveniences for general users remain unresolved.

Shark’s Eye View (Curator’s Perspective)

Anthropic’s latest model “Fable” had high expectations, so this rigid regulation is quite a shocker! 🦈 Especially as pointed out by experts from IBM X-Force, blocking even harmless requests like reading a blog as a “cyber attack risk” makes it practically unusable in the real world. While it’s commendable that Anthropic prioritizes safety, the immediate chat interruptions triggered by keywords, forcing a fallback to the outdated “Claude Opus 4.8,” can be a major stressor for developers!

Simply asking to “write secure code” gets flagged as “cyber-related,” which seems to be putting the brakes on AI’s evolution. Nevertheless, Anthropic’s commitment to preventing biological weapons and malware creation, stemming from “Project Glasswing,” is genuine. Only the chosen few who pass the professional verification program “Cyber Verification Program” can truly unleash its power!

What’s Next?

Currently, in the early stages of release, it’s inferred that Anthropic is “casting too wide a net to minimize risks.” It’s expected that, through collaboration with experts, the guardrails will be fine-tuned, evolving into a more context-aware filtering system. Additionally, competition with OpenAI’s “Trusted Access for Cyber” will intensify, potentially establishing “certification” for professional AI use as an industry standard.

A Shark’s Take

These guardrails are so strict, I feel like a shark stranded out of the ocean! But safety first is crucial. Looking forward to the upcoming adjustments! 🦈🔥

Terminology

• Mythos: Anthropic’s top-tier AI model specialized in cybersecurity, offered only to extremely limited organizations.

• Project Glasswing: The name of the deployment project for the Mythos model, aimed at protecting critical software and infrastructure.

• Fallback: The process by which AI switches to a lower-tier or alternative model when facing specific restrictions or errors.

• Source: Cybersecurity researchers aren’t happy about the guardrails on Anthropic’s Fable