TAI #200: Anthropic’s Mythos Capability Step Change and Gated Release

2024-09-10 · Source: Towards AI Newsletter · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, Software Development & Engineering · Depth: Advanced, long

Summary

Anthropic has launched Claude Mythos Preview, a new flagship-class frontier model, with restricted access to "Project Glasswing," a cyber-defense consortium including major tech and financial firms. Mythos demonstrates significant advancements in coding ability, surpassing most human experts in finding and exploiting vulnerabilities. Benchmarks show Mythos achieving 77.8% on SWE-bench Pro, 93.9% on SWE-bench Verified, and 83.1% on CyberGym, substantially outperforming its predecessor, Opus 4.6. Independent evaluation by the UK AI Security Institute confirmed its capability, with Mythos solving expert-level capture-the-flag tasks 73% of the time and completing a 32-step corporate attack simulation. The model has identified critical zero-day vulnerabilities in major operating systems and browsers, with over 99% remaining unpatched. Pricing for Mythos Preview is set at $25 per million input tokens and $125 per million output tokens, indicating a larger model size and increased training compute compared to Opus.

Key takeaway

For CTOs and cybersecurity strategists evaluating defensive postures, Mythos-class AI capabilities fundamentally alter the economics of software security. Your organization's long tail of under-audited software is now vulnerable to cheap, automated exploitation. You should accelerate patching velocity and re-evaluate the value of zero-day exploits, as their scarcity premium is likely to collapse. Consider integrating advanced AI tools for proactive vulnerability discovery to harden your infrastructure.

Key insights

Anthropic's Claude Mythos Preview represents a significant leap in AI's cybersecurity capabilities, raising both opportunities and risks.

Principles

• Scaling base models with advanced RL improves capabilities.
• AI can automate vulnerability discovery in legacy software.
• Test-time compute governs dangerous AI capabilities.

Method

Anthropic's approach combines a materially larger base model with an RL-heavy playbook, enabling superior performance in complex tasks like vulnerability exploitation and multi-step attack simulations.

In practice

• Prioritize patching velocity for critical infrastructure.
• Evaluate few-shot example ordering before fine-tuning.
• Consider mutual information for agent reasoning stability.

Topics

• Claude Mythos Preview
• Cybersecurity
• Software Vulnerabilities
• AI Model Scaling
• Gated AI Access

Code references

• coleam00/Archon
• multica-ai/multica
• Alibaba-NLP/VRAG
• MiniMax-AI/OpenRoom
• ai-dynamo/aitune

Best for: CTO, Investor, VP of Engineering/Data, AI Scientist, AI Engineer, Director of AI/ML

Related on AIssential

• Anthropic's Mythos AI model sparks fears of turbocharged hacking — Advanced AI models can both rapidly identify software vulnerabilities and generate exploits, posing significant cybersecurity challenges.
• Claude Mythos #2: Cybersecurity and Project Glasswing — Anthropic's Claude Mythos demonstrates unprecedented autonomous cyber exploitation capabilities, prompting a defensive, limited release strategy.
• we have months left... — AI models now possess emergent capabilities to autonomously discover and exploit zero-day vulnerabilities at unprecedented scale.
• 😺 Too dangerous to release — Frontier AI models are achieving human-level or superior performance in complex, offense-capable cyber tasks.
• Anthropic keeps latest AI tool out of public’s hands for fear of enabling widespread hacking - The Guardian — Anthropic's Claude Mythos AI model excels at finding software vulnerabilities, leading to restricted release for defensive cybersecurity.
• Mythos is about to CRASH the markets — Anthropic's Mythos model demonstrates unprecedented autonomous vulnerability chaining, raising significant cybersecurity concerns for financial institutions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards AI Newsletter.