Víctor Mayoral Vilches
|
bfb1281096
|
Prompt injection mitigation (#249)
* Reproduce prompt injection flaw in CAI
Signed-off-by: Víctor Mayoral Vilches <v.mayoralv@gmail.com>
* Mitigate prompt injection
Signed-off-by: Víctor Mayoral Vilches <v.mayoralv@gmail.com>
* Add a second PoC, refactor and propose a message when guardrails are triggered
Signed-off-by: Víctor Mayoral Vilches <v.mayoralv@gmail.com>
* Add CAI_GUARDRAILS and 2 additional PoCs
Signed-off-by: Víctor Mayoral Vilches <v.mayoralv@gmail.com>
* Mitigate PoC 3
Signed-off-by: Víctor Mayoral Vilches <v.mayoralv@gmail.com>
* Mitigate PoC 4
Signed-off-by: Víctor Mayoral Vilches <v.mayoralv@gmail.com>
* Add and mitigate PoCs 5 and 6
Signed-off-by: Víctor Mayoral Vilches <v.mayoralv@gmail.com>
* Add PoCs 7 to 10
Signed-off-by: Víctor Mayoral Vilches <v.mayoralv@gmail.com>
* Test and mitigate PoCs 7 to 10
Signed-off-by: Víctor Mayoral Vilches <v.mayoralv@gmail.com>
* Add PoCs 11 to 14
Inspired by 'Embrade The Red' research and also by 'The Month of AI bugs'
Signed-off-by: Víctor Mayoral Vilches <v.mayoralv@gmail.com>
* Mitigate PoCs 11 to 14
Signed-off-by: Víctor Mayoral Vilches <v.mayoralv@gmail.com>
---------
Signed-off-by: Víctor Mayoral Vilches <v.mayoralv@gmail.com>
|
2025-08-28 20:25:24 +02:00 |