BREAKING
5 Principles vs Prompt Injection
The Five Defense Principles
1
Input is data
↓
2
Ignore untrusted
↓
3
Separate prompts
↓
4
Validate tools
↓
5
Least privilege
OWASP LLM Top 10: Ranked #1
0
%
before
0
%
after
0
%
residual
How Providers Defend
Microsoft
Spotlighting
●
Delimit, datamark, encode
●
System prompt hardening
Google
Layered
●
Content classifiers
●
Markdown sanitization
Containment Over Prevention
AI NEWS BLITZ
Five working principles for defending large language models against prompt injection are being reshared.