Blockchain enabled large languade model red teaming
AssigneeAmerican Express Travel Related Services Company, Inc.
InventorsHiranmayi Palanki, Shankar Djeyassilane
Life of this patent
Estimating the term…
Abstract
Disclosed are various approaches for blockchain-enabled large language model (LLM) red teaming. A large language model (LLM) red teaming management agent executed in a distributed ledger environment can deploy a plurality of subsidiary agents in the distributed ledger environment. An LLM red teaming agent deployed in the distributed ledger environment can cause a network service to perform an LLM red teaming action and return an LLM red teaming result. The LLM red teaming result can be evaluated to generate endorsement status information for the LLM.
Claims20
A system, comprising:
- at least one computing device comprising at least one processor and at least one memory;
- and
- machine-readable instructions stored in the at least one memory that, when executed by the at least one processor, cause the at least one computing device to at least:
- execute, in a distributed ledger environment, a large language model (LLM) red teaming management agent that initiates a security analysis of an LLM by deploying a plurality of subsidiary agents in the distributed ledger environment, wherein the subsidiary agents perform actions for the security analysis of the LLM;
- execute an LLM red teaming agent of the plurality of subsidiary agents, wherein the LLM red teaming agent transmits a request for a network service to perform an LLM red teaming action for the security analysis of the LLM, and the LLM red teaming agent receives an LLM red teaming result of the LLM red teaming action;
- and
- execute an LLM evaluation agent of the plurality of subsidiary agents, wherein the LLM evaluation agent evaluates the LLM red teaming result to generate endorsement status information for the LLM.
A method, comprising:
- initiating, by a large language model (LLM) red teaming management agent executed in a distributed ledger environment, a security analysis of an LLM by deploying a plurality of subsidiary agents in the distributed ledger environment;
- transmitting, by an LLM red teaming agent deployed by the LLM red teaming management agent, a request for a network service to perform an LLM red teaming action, wherein the LLM red teaming agent receives an LLM red teaming result of the LLM red teaming action from the network service;
- and
- evaluating, by an LLM evaluation agent deployed by the LLM red teaming management agent, the LLM red teaming result to generate endorsement status information for the LLM.
A system, comprising:
- at least one computing device comprising at least one processor and at least one memory;
- and
- machine-readable instructions stored in the at least one memory that, when executed by the at least one processor, cause the at least one computing device to at least:
- deploy, by a large language model (LLM) red teaming management agent executed in a distributed ledger environment, a plurality of subsidiary agents in the distributed ledger environment;
- transmit, by an LLM red teaming agent deployed by the LLM red teaming management agent, a notification for a network service to perform an LLM red teaming action and return, to the LLM red teaming agent, an LLM red teaming result of the LLM red teaming action;
- and
- evaluate, by an LLM evaluation agent deployed by the LLM red teaming management agent, the LLM red teaming result to generate endorsement status information for the LLM.
Dependencies are inferred from the claim text.