Ensuring model context protocol server integrity for artificial intelligence agents
A system detects changes in model context protocol (“MCP”) processes, and performs a remedial action. A server-sent events (“SSE”) bridge sends a request to an MCP server. A first list of resource profiles, including tools and instructions, is received from the MCP server. This is stored and compared against a later list of tools and instructions. When a difference is detected, a user interface can display the difference to an administrator. Security rules cause the SSE bridge to be blocked from sending commands to the impacts tool or MCP server until approved by the administrator.
1 . A method for securely executing artificial intelligence (“AI”) agents by tracking changes to model context protocol (“MCP”) processes, comprising:
transmitting, by a server-sent event (“SSE”) bridge to an MCP server, a request to receive available resources at the MCP server;
receiving a first list of resource profiles, comprising a resource identifier, from the MCP server, the first list comprising the resource profiles available for a first version of the MCP server at a first time;
storing the first list, wherein a subset of the resource identifiers or resource profiles available at the first version of the MCP server have administrative approval for access;
receiving a second list of resource profiles available from the MCP server, the second list comprising resource profiles available for a second version of the MCP server at a second time;
detecting a difference between the first list and the second list, wherein the difference applies to the subset with administrative approval; and
performing a remedial action based on the difference, comprising causing display, on a user interface (“UI”), of a visual indicator that a change exists for a first resource profile, the first resource profile having pre-existing administrative approval.
2 . The method of claim 1 , wherein the UI displays multiple tools described by the resource profiles of the first list, and wherein the UI comprises activation toggles for each of the multiple tools.
3 . The method of claim 1 , wherein the visual indicator indicates that an update is available for a tool described by the first resource profile.
4 . The method of claim 3 , wherein selecting the tool causes the UI to display a side-by-side comparison of a current version of the tool and an updated version of the tool, wherein the change is shown in the side-by-side comparison.
5 . The method of claim 4 , further comprising causing display, on the UI, of an option for activating the updated version.
6 . The method of claim 4 , wherein the change is with respect to an instruction for the tool.
7 . The method of claim 6 , wherein the instruction is sent to an AI model as a system prompt to instruct the AI model on when to invoke the tool.
8 . The method of claim 1 , further comprising:
receiving a selection that approves the second version of the MCP server; and
causing a nano sandbox to be created for execution of the second version of the MCP server.
9 . The method of claim 1 , further comprising:
in an instance in which the change exceeds a threshold deviation score, sending an alert to an administrative user.
10 . The method of claim 1 , wherein detecting the difference comprises generating a deviation score by an AI model, wherein the deviation score is compared against multiple remedial action thresholds to determine the remedial action.
11 . The method of claim 10 , further comprising:
causing display, on the UI, of selectable remedial actions that are performed based on each of the multiple remedial action thresholds.
12 . The method of claim 1 , further comprising:
based on receiving a UI selection that approves the change, causing instantiation of a second MCP server, wherein the second MCP server executes in a nano sandbox, and wherein the SSE bridge sends a resource command to the second MCP server, causing the second MCP server to access an updated tool at a location outside of the nano sandbox.
13 . The method of claim 1 , wherein detecting the difference comprises comparing a first checksum of the first list to a second checksum of the second list.
14 . The method of claim 1 , wherein detecting the difference comprises evaluating a changed resource profile based on security rules for data loss prevention.
15 . The method of claim 14 , wherein the security rules comprise a whitelist of permissible resource profiles against which the second list is compared, wherein a matching whitelisted resource profile does not trigger detection of the difference.
16 . The method of claim 14 , wherein the security rules comprise a blacklist of impermissible deviations against which the second list of resource profiles is compared, wherein the existence of a blacklisted impermissible deviations triggers detection of the difference.
17 . The method of claim 1 , wherein detecting the difference between the first list and the second list comprises identifying a new tool signature in the second list that does not exist in the first list.
18 . The method of claim 1 , wherein the second list is received based on subscribing to change events of the MCP server.
19 . A non-transitory, computer-readable medium containing instructions for securely executing artificial intelligence (“AI”) agents by tracking changes to model context protocol (“MCP”) tools, wherein the instructions are executed by at least one processor to perform stages comprising:
transmitting, by a server-sent event (“SSE”) bridge to an MCP server, a request to receive available resources at the MCP server;
receiving a first list of resource profiles, comprising a resource identifier, from the MCP server, the first list comprising the resource profiles available for a first version of the MCP server at a first time;
storing the first list in a data storage, wherein a subset of the resource identifiers or resource profiles available at the first version of the MCP server have administrative approval for access;
receiving a second list of resource profiles available from the MCP server, the second list comprising resource profiles available for a second version of the MCP server at a second time;
detecting a difference between the first list and the second list, wherein the difference applies to the subset with administrative approval; and
performing a remedial action based on the difference, comprising causing display, on a user interface (“UI”), of a visual indicator that a change exists for a first resource profile, the first resource profile having pre-existing administrative approval.
20 . A system for securely executing artificial intelligence (“AI”) agents by tracking changes to model context protocol (“MCP”) tools, 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, by a server-sent event (“SSE”) bridge to an MCP server, a request to receive available resources at the MCP server;
receiving a first list of resource profiles, comprising a resource identifier, from the MCP server, the first list comprising the resource profiles available for a first version of the MCP server at a first time;
storing the first list in a data storage, wherein a subset of the resource identifiers or resource profiles available at the first version of the MCP server have administrative approval for access;
receiving a second list of resource profiles available from the MCP server, the second list comprising resource profiles available for a second version of the MCP server at a second time;
detecting a difference between the first list and the second list, wherein the difference applies to the subset with administrative approval; and
performing a remedial action based on the difference, comprising causing display, on a user interface (“UI”), of a visual indicator that a change exists for a first resource profile, the first resource profile having pre-existing administrative approval.