Anthropic LLM Models

Anthropic LLM Models

  • Context:

  • Anthropic, the San Francisco-based AI company behind the Claude family of Large Language Models (LLMs), recently announced its most powerful model to date:

  • Mythos.

  • Recognising the immense cybersecurity risks of such advanced AI falling into the wrong hands, the company has decided against releasing the model widely to the general public.

  • Project Glasswing:

  • Instead of a public release, Mythos is restricted to a consortium of over 40 companies under an initiative known as Project Glasswing.

  • The core objective is to allow the developers of foundational IT systems to get a crucial "head start."

  • They can utilize Mythos to scan decades-old code and identify previously undetected software bugs and vulnerabilities before malicious actors can use similar Mythos-class models to launch automated cyberattacks.

  • Implications for India:

  • Vulnerability of Bespoke Software:

  • While the Indian IT industry will ultimately benefit from global platforms patching their software via Glasswing, India often builds its own bespoke software solutions.

  • These proprietary systems could remain highly vulnerable if attacked by sophisticated, autonomous AI models.

  • Regulatory Response:

  • No Indian IT firm is currently publicly listed as a Glasswing partner.

  • Consequently, the Data Security Council of India (under Nasscom) and the Indian Computer Emergency Response Team (CERT-In) under the IT Ministry are actively convening to study the national security implications of this exclusion.

  • The Claude LLM Hierarchy:

  • Anthropic's LLMs are highly regarded for their coding, productivity, and reasoning capabilities, typically running through command-line interfaces or dedicated apps.

  • In increasing order of sophistication, the models are structured as follows:

Claude Model

Relative Sophistication

Accessibility & Capability Profile

Haiku

Base level

Publicly available.

Noted for strong performance, reasoning, and the ability to compose code.

Sonnet

Intermediate level

Publicly available.

Shares the core performance and reasoning capabilities of the Claude family.

Opus

Advanced level

Publicly available.

Highly sophisticated, capable of performing complex tasks agentically.

Mythos

Most Powerful

Restricted access.

Uniquely capable of identifying deeply hidden vulnerabilities in legacy code.