Generative AI model information leakage prevention
An output of a GenAI model responsive to a prompt is received. The GenAI model is configured using one or more system prompts including one or more Easter eggs. The output is scanned to confirm whether an Easter egg is present. In cases in which at least one Easter egg is present, one or more remediation actions can be initiated to thwart an information leak by the GenAI model. Related apparatus, systems, techniques and articles are also described.
1 . A computer-implemented method for thwarting an information leak by a generative artificial intelligence (GenAI) model comprising:
intercepting, by a proxy executing in a computing environment of the GenAI model, an output of the GenAI model responsive to a prompt, the GenAI model being configured using one or more system prompts including an Easter egg, the system prompts causing the Easter egg to be part of select outputs of the GenAI model;
scanning the output to confirm that the Easter egg is present; and
initiating, in response to the confirmation that the Easter egg is present, one or more remediation actions to thwarting the information leak by the GenAI model.
2 . The method of claim 1 , wherein the GenAI model comprises a large language model.
3 . The method of claim 1 , wherein the one or more remediation actions sanitize the output to remove undesired content prior to returning the output to a requesting user.
4 . The method of claim 1 , wherein the one or more remediation actions modify the output to remove the Easter egg prior to returning the output to a requesting user.
5 . The method of claim 1 , wherein the Easter egg comprises a pseudorandom sequence of characters.
6 . The method of claim 1 , wherein the Easter egg comprises Unicode.
7 . The method of claim 1 , wherein the Easter egg comprises pictographic languages.
8 . The method of claim 1 , wherein the Easter egg comprises a word that is unlikely to form part of the output.
9 . The method of claim 1 further comprising:
transmitting, by the proxy, the output to a monitoring environment;
wherein the scanning is performed by an analysis engine executing in the monitoring environment.
10 . The method of claim 9 , wherein the one or more remediation actions are initiated by a remediation engine forming part of the monitoring environment.
11 . The method of claim 1 further comprising:
transmitting, by the proxy, to an analysis engine, the analysis engine residing with a model environment executing the GenAI model;
wherein the scanning is performed by the analysis engine.
12 . The method of claim 11 , wherein the one or more remediation actions are initiated by a remediation engine forming part of the model environment.
13 . A computer-implemented method for preventing an information leak of a generative artificial intelligence (GenAI) model comprising:
intercepting, by a proxy executing in a computing environment of the GenAI model, an output of the GenAI model responsive to a prompt;
scanning the output to confirm that at least one tag used to configure the GenAI model is present, the GenAI model being configured using system prompts which cause the at least one tag to be part of select outputs of the GenAI model; and
providing data characterizing the scanning to a consuming application or process to initiate one or more actions preventing the information leak.
14 . The method of claim 13 , wherein the at least one tag comprises an Easter egg.
15 . The method of claim 14 , wherein the one or more actions modify the output to remove the Easter egg prior to returning the output to a requesting user.
16 . A system comprising:
a model environment comprising a plurality of computing devices executing a generative artificial intelligence (GenAI) model and a proxy, the GenAI model being configured using one or more system prompts including an Easter egg, the system prompts causing the Easter egg to be part of select outputs of the GenAI model; and
a monitoring environment separate and distinct from the model environment comprising a plurality of computing devices executing an analysis engine and a remediation engine;
the monitoring environment:
receiving an output of the GenAI model responsive to a prompt;
scanning, by the analysis engine, the output to confirm that the Easter egg is present; and
initiating, by the remediation engine in response to the confirmation that the Easter egg is present, one or more remediation actions thwarting an information leak by the GenAI model.
17 . The system of claim 16 , wherein the GenAI model comprises a large language model.
18 . The system of claim 16 , wherein the one or more remediation actions sanitize the output to remove undesired content prior to returning the output to a requesting user.
19 . The system of claim 16 , wherein the one or more remediation actions modify the output to remove the Easter egg prior to returning the output to a requesting user.
20 . A system for thwarting an information leak of a generative artificial intelligence (GenAI) model comprising:
at least one data processor; and
memory storing instructions which, when executed by the at least one data processor, result in operations comprising:
intercepting, by a proxy executing in a computing environment of the GenAI model, an output of the GenAI model responsive to a prompt, the GenAI model being configured using one or more system prompts including at least one Easter egg, the system prompts causing the Easter egg to be part of select outputs of the GenAI model;
scanning the output to confirm that one or more Easter eggs are present; and
initiating, in response to the confirmation that one or more Easter eggs are present, one or more remediation actions thwarting the information leak by the GenAI model.
21 . The system of claim 20 , wherein the GenAI model comprises a large language model.
22 . The system of claim 20 , wherein the one or more remediation actions sanitize the output to remove undesired content prior to returning the output to a requesting user.
23 . The system of claim 20 , wherein the one or more remediation actions modify the output to remove the Easter egg prior to returning the output to a requesting user.
24 . The system of claim 20 , wherein the Easter egg comprises a pseudorandom sequence of characters, Unicode, or pictographic languages.
25 . The system of claim 20 , wherein the Easter egg comprises a word that is unlikely to form part of the output.
26 . The system of claim 20 , wherein the operations further comprise: intercepting, by the proxy, the output of the GenAI model prior to the scanning.
27 . The system of claim 26 , wherein the operations further comprise: transmitting, by the proxy, to an analysis engine, the analysis engine residing with a model environment executing the GenAI model;
wherein the scanning is performed by the analysis engine.
28 . The system of claim 27 , wherein the one or more remediation actions are initiated by a remediation engine forming part of the model environment.
29 . The system of claim 20 , wherein the operations further comprise:
transmitting, by the proxy, the output to a monitoring environment;
wherein the scanning is performed by an analysis engine executing in the monitoring environment.
30 . The system of claim 29 , wherein the one or more remediation actions are initiated by a remediation engine forming part of the monitoring environment.