mirror of
https://github.com/splunk/DECEIVE.git
synced 2025-07-01 08:37:28 -04:00
Added DECEIVE image to README
This commit is contained in:
BIN
DECEIVE.png
Normal file
BIN
DECEIVE.png
Normal file
Binary file not shown.
After Width: | Height: | Size: 600 KiB |
@ -1,5 +1,7 @@
|
|||||||
# DECEIVE
|
# DECEIVE
|
||||||
|
|
||||||
|
<img align="right" src="DECEIVE.png" alt="A cybercriminal interacts with a ghostly, AI-driven honeypot system">
|
||||||
|
|
||||||
DECEIVE, the **DECeption with Evaluative Integrated Validation Engine**, is a high-interaction, low-effort honeypot system. Unlike most high-interaction honeypots, DECEIVE doesn't provide attackers with access to any actual system. AI actually does all the work of simulating a realistic honeypot system based on a configurable system prompt that describes what type of system you want to simulate. Unlike many other high-interaction honeypots which require substantial effort to seed with realistic users, data, and applications, DECEIVE's AI backend will do all this for you, automatically.
|
DECEIVE, the **DECeption with Evaluative Integrated Validation Engine**, is a high-interaction, low-effort honeypot system. Unlike most high-interaction honeypots, DECEIVE doesn't provide attackers with access to any actual system. AI actually does all the work of simulating a realistic honeypot system based on a configurable system prompt that describes what type of system you want to simulate. Unlike many other high-interaction honeypots which require substantial effort to seed with realistic users, data, and applications, DECEIVE's AI backend will do all this for you, automatically.
|
||||||
|
|
||||||
This version of DECEIVE simulates a Linux server via the SSH protocol. It will log all the user inputs, the outputs returned by the LLM backend, as well as a summary of each session after they end. It'll even tell you if it thinks a users' session was benign, suspicious, or outright malicious.
|
This version of DECEIVE simulates a Linux server via the SSH protocol. It will log all the user inputs, the outputs returned by the LLM backend, as well as a summary of each session after they end. It'll even tell you if it thinks a users' session was benign, suspicious, or outright malicious.
|
||||||
|
Reference in New Issue
Block a user