BREAKING
Anthropic Drafts Jailbreak Severity Rules
Four Partners At The Table
Co-drafters
joint
●
Anthropic
●
Amazon
●
Microsoft
●
Google
Project Glasswing
open
●
Cyber defense initiative
●
Members invited to join
●
Announced April 7, 2026
Why The Framework Now
1
Export limits June 12
↓
2
Amazon jailbreak report
↓
3
New safety classifier
↓
4
Models redeployed
0
%
Technique blocked
0
%
Mythos on CyberGym
0
%
Opus 4.6
Four Assessment Axes
Capability
axes
●
Capability gain
●
Breadth of gain
Exploit
axes
●
Ease of weaponization
●
Discoverability
Consensus Is The Next Test
AI NEWS BLITZ
Anthropic is building an industry framework to grade AI jailbreak severity.