INDEX

Explanations

data exfiltration or command and control

New Auto-Interp

Top Features by Cosine Similarity

Configuration

Prompts (Dashboard)

10,000 prompts, 128 tokens each

Dataset (Dashboard)

lmsys/lmsys-chat-1m

Embeds

IFrame

Link

Not in Any Lists

No Comments

Negative Logits

 Vulner

-0.10

 safety

-0.10

 safe

-0.09

 robust

-0.09

 offending

-0.09

 vulnerable

-0.09

 Grove

-0.09

 martial

-0.09

 infr

-0.09

POSITIVE LOGITS

 lateral

0.16

 privileges

0.13

 employee

0.13

 compromise

0.12

 privilege

0.12

Bre

0.11

 Insider

0.11

 comprom

0.11

BY

0.11

 privileged

0.11

Activations Density 0.033%