AI capabilities are accelerating faster than most realize Anthropic’s Frontier Red Team —

AI capabilities are accelerating faster than most realize

Anthropic’s Frontier Red Team — the group tasked with stress-testing AI models before they go public — has been tracking some striking trends:

📈 Cyber offense skills are surging
Claude’s ability to solve complex coding & penetration testing challenges jumped from 2% in 2023 to 49% in 2024.

🧬 Expert-level scientific knowledge is emerging
In biology and chemistry, AI accuracy is now approaching human expert levels in some areas — a dual-use capability with obvious safety implications.

🔐 Undergraduate-level cybersecurity + expert biology
Models can now combine offensive cyber skills with deep scientific understanding, a pairing that could be highly valuable — or highly dangerous.

The takeaway:
AI is moving from a support tool to a capable operator in sensitive domains. If we don’t test, measure, and gate these abilities, we risk being caught off-guard.

Anthropic’s proactive approach — using real-world hacking contests, domain-expert evaluations, and automated red-teaming — is exactly the mindset we need.

💡 The question isn’t just what AI can do today… it’s what it will be able to do tomorrow — and whether we’re ready.