Protected remote distribution of artificial intelligence agents using secure tunnels
View Patent ↗Systems are disclosed for securely deploying artificial intelligence (“AI”) agents to remote device environments over secure connections. A cloud connector delivered to the remote device authenticates to a platform connection manager, which opens a first secure tunnel without the need for opening additional ports at the remote device. The AI platform receives a request to execute an AI agent and determines a compliance status for the remote device by applying security rules. Based on the status and the request, the platform modifies a manifest file of the AI agent to identify local agent objects (e.g., an AI model, vector database, or dataset) and queries the device to confirm access. The AI platform then transmits the manifest to the remote device, where an agent executor runs the agent according to the manifest. Execution records are returned over the secure tunnel to provide centralized oversight and auditability.
1 . A method for deploying artificial intelligence (“AI”) agents over secure cloud connections, comprising:
transmitting a cloud connector installer to a remote device, the cloud connector installer being transmitted from an AI platform, the cloud connector installer causing a cloud connector to be installed at the remote device;
receiving, from the cloud connector executing at the remote device, an authentication request;
in response to the authentication request, opening a first secure tunnel between the cloud connector and a connection manager of the AI platform, wherein opening the first secure tunnel comprises receiving, at the connection manager, a socket connection originating from the cloud connector at the remote device, and wherein the first secure tunnel is established exclusively through an outbound connection from the remote device, the first secure tunnel multiplexing, over the socket connection, a plurality of substreams comprising a control-plane substream, a data-plane substream, and a telemetry substream;
receiving, from the remote device over the first secure tunnel, a request to execute an AI agent at the remote device;
determining a compliance status of the remote device, wherein the compliance status is determined by applying security rules for local execution of the AI agent;
based on the compliance status and the request, modifying a manifest file of the AI agent, wherein the modified manifest file identifies an agent object within a network of the remote device, the agent object comprising at least one of an AI model, a vector database, and a dataset;
querying the remote device over the secure tunnel to determine whether the remote device has access to the agent object;
transmitting the manifest file over the secure tunnel to the remote device, wherein an agent executor executes the AI agent according to the manifest file; and
receiving an execution record of the AI agent over the secure tunnel.
2 . The method of claim 1 , further comprising, based on the querying, determining whether to send the agent object or an object configuration to the remote device, wherein the object configuration allows the remote device to interact with the agent object, the object configuration comprising an endpoint and credentials for the agent object.
3 . The method of claim 1 , further comprising transmitting an agent executor to the remote device, the agent executor being configured to execute the AI agent according to the manifest file.
4 . The method of claim 1 , further comprising:
determining the compliance status of the remote device has changed; and
making an additional modification to the manifest file based on the changed compliance status.
5 . The method of claim 1 , comprising:
receiving data over the secure tunnel, the data being an input to or an output from the agent object;
modifying the data with guardrails; and
transmitting, over the secure tunnel, the modified data to the cloud connector, wherein the modified data invokes further execution of the AI agent.
6 . The method of claim 5 , wherein the guardrails are configured to detect in the data at least one of prompt injection and personally identifiable information (“PII”).
7 . The method of claim 1 , further comprising:
transmitting, over the secure tunnel, a management profile to the remote device, wherein a conditional operation of an agent executor is invoked based on compliance with the management profile by the remote device or a user device interacting with the AI agent.
8 . The method of claim 7 , wherein the management profile defines which of the security rules are invoked by the agent executor.
9 . The method of claim 1 , wherein the manifest is modified by prioritizing the agent object relative to a second agent object at the AI platform.
10 . The method of claim 1 , wherein transmitting the cloud connector installer comprises transmitting a command to the remote device, wherein the command instructs the remote device on where to access the cloud connector installer.
11 . The method of claim 10 , wherein the command is sent to a docker environment, causing the docker environment to fetch the cloud connector installer.
12 . The method of claim 1 , further comprising configuring the AI agent, wherein configuring the AI agent comprises:
receiving, over the secure tunnel, identification of multiple agent objects that are within the network of the remote device;
invoking display of a user interface (“UI”) that comprises options to select and connect one or more of the multiple agent objects; and
receiving a selection to add the agent object to the AI agent.
13 . The method of claim 12 , wherein the UI comprises options to select and connect additional agent objects that are outside of the network of the remote device.
14 . The method of claim 13 , wherein a second agent object is added to the AI agent, and wherein the second agent object is accessed or executed by the AI platform.
15 . The method of claim 12 , wherein the manifest file is automatically generated based on the selected and connected one or more of the multiple agent objects.
16 . The method of claim 1 , wherein deployment of the AI agent is contingent on prior simulated execution of the AI agent at the AI platform, and wherein simulating execution comprises:
invoking display of a user interface (“UI”) that receives inputs for the AI agent; and
invoking display of outputs for the AI agent based on the received inputs.
17 . The method of claim 1 , further comprising logging the request from the agent executor.
18 . The method of claim 17 , further comprising logging context of the request, the context identifying a user and the AI agent, wherein the context is received over the secure tunnel from the cloud connector.
19 . A non-transitory, computer-readable medium containing instructions for deploying artificial intelligence (“AI”) agents over secure cloud connections, wherein the instructions are executed by at least one processor to perform stages comprising:
transmitting a cloud connector installer to a remote device, the cloud connector installer being transmitted from an AI platform, the cloud connector installer causing a cloud connector to be installed at the remote device;
receiving, from the cloud connector executing at the remote device, an authentication request;
in response to the authentication request, opening a first secure tunnel between the cloud connector and a connection manager of the AI platform, wherein opening the first secure tunnel comprises receiving, at the connection manager, a socket connection originating from the cloud connector at the remote device, and wherein the first secure tunnel is established exclusively through an outbound connection from the remote device, the first secure tunnel multiplexing, over the socket connection, a plurality of substreams comprising a control-plane substream, a data-plane substream, and a telemetry substream;
receiving, from the remote device over the first secure tunnel, a request to execute an AI agent at the remote device;
determining a compliance status of the remote device, wherein the compliance status is determined by applying security rules for local execution of the AI agent;
based on the compliance status and the request, modifying a manifest file of the AI agent, wherein the modified manifest file identifies an agent object within a network of the remote device, the agent object comprising at least one of an AI model, a vector database, and a dataset;
querying the remote device over the secure tunnel to determine whether the remote device has access to the agent object;
transmitting the manifest file over the secure tunnel to the remote device, wherein an agent executor executes the AI agent according to the manifest file; and
receiving an execution record of the AI agent over the secure tunnel.
20 . A system for deploying artificial intelligence (“AI”) agents over secure cloud connections, the system comprising:
at least one physical non-transitory, computer-readable medium comprising instructions; and
at least one processor that executes the instructions to perform stages comprising:
transmitting a cloud connector installer to a remote device, the cloud connector installer being transmitted from an AI platform, the cloud connector installer causing a cloud connector to be installed at the remote device;
receiving, from the cloud connector executing at the remote device, an authentication request;
in response to the authentication request, opening a first secure tunnel between the cloud connector and a connection manager of the AI platform, wherein opening the first secure tunnel comprises receiving, at the connection manager, a socket connection originating from the cloud connector at the remote device, and wherein the first secure tunnel is established exclusively through an outbound connection from the remote device, the first secure tunnel multiplexing, over the socket connection, a plurality of substreams comprising a control-plane substream, a data-plane substream, and a telemetry substream;
receiving, from the remote device over the first secure tunnel, a request to execute an AI agent at the remote device;
determining a compliance status of the remote device, wherein the compliance status is determined by applying security rules for local execution of the AI agent;
based on the compliance status and the request, modifying a manifest file of the AI agent, wherein the modified manifest file identifies an agent object within a network of the remote device, the agent object comprising at least one of an AI model, a vector database, and a dataset;
querying the remote device over the secure tunnel to determine whether the remote device has access to the agent object;
transmitting the manifest file over the secure tunnel to the remote device, wherein an agent executor executes the AI agent according to the manifest file; and
receiving an execution record of the AI agent over the secure tunnel.