| MCP01Token Mismanagement & Secret Exposure | Hard-coded credentials, long-lived tokens, and secrets stored in model memory or protocol logs. Prompt injection, a compromised context, or a debug trace can hand them to an attacker. | Keep tokens out of agent config, and never leave an agent a standing credential. Umbra issues short-lived leases that Alyria Beacon injects at the moment of use and discards afterward. Beacon flags plaintext credentials already sitting in mcpServers env blocks across the fleet. |
| MCP02Privilege Escalation via Scope Creep | Loosely defined or temporary permissions expand over time: an agent that started with read access ends up able to modify repositories, control systems, or exfiltrate data. | Grant a capability, not a standing grant. Lyra policy is capability-brokered: agent X may call tool Y against resource Z for this task, and nothing more. OAuth 2.1 on remote servers, as the MCP spec directs, covers the transport; the action-level scope is still enforced where the call is made. |
| MCP03Tool Poisoning | Malicious, misleading, or biased instructions hidden in tool descriptions, schemas, or results steer the model without anyone noticing. | Pin the version you vetted and inspect what the server says before the agent acts on it. Beacon inventories every configured MCP server and flags unpinned entries; a content-safety scan runs over tool responses before they reach the model. |
| MCP04Software Supply Chain Attacks & Dependency Tampering | A compromised package, backdoored server, or poisoned dependency alters agent behavior or introduces an execution-level backdoor. | Pin exact versions and verify provenance before install. An unqualified entry resolves whatever copy is in the cache at launch, so two machines can run different code from the same config line. Beacon tracks the version each endpoint actually runs and can refuse an off-list server at launch once enforcement is on. |
| MCP05Command Injection & Execution | An agent constructs and executes shell commands, scripts, or code from untrusted input without validation or sanitization. | Deny the call before it runs, and bound what any spawned process can do. Codex's sandbox and approval policy and Claude Code's sandboxing and PreToolUse hooks are the vendors' own controls; user-space runtime observation covers processes that bypass the mediated path. |
| MCP06Prompt Injection via Contextual Payloads | Instructions hidden in pages, issues, or tool output redirect the agent's own reasoning. The model is the interpreter and text is the payload. | Treat everything the agent reads as untrusted, and write the follow-through down. Content-safety scans run on tool responses and shared-memory recall, and sequence policy closes the second step: if the session has read a flagged payload, block network egress. |
| MCP07Insufficient Authentication & Authorization | MCP servers, tools, or agents fail to verify identities or enforce access controls, so callers reach data and actions they were never entitled to. | Bind the agent to an identity your IdP knows and enforce per-identity grants in policy. HTTP servers can require OAuth 2.1 as the MCP spec directs; a stdio server runs as the local user with no identity boundary at all, so the grant that matters is the one evaluated at the call. |
| MCP08Lack of Audit and Telemetry | Limited telemetry from MCP servers and agents impedes investigation and incident response; nobody can reconstruct what an agent actually did. | Record every tool call, the policy decision attached to it, and the identity that made it, in one signed, hash-chained chain, then route the signal into the SIEM over OpenTelemetry so existing detections see agent activity in the same pipeline as everything else. |
| MCP09Shadow MCP Servers | Unapproved MCP deployments operating outside security governance, often with default credentials, permissive configuration, or unsecured APIs. | Keep a live inventory of the config each client holds, including stdio servers that never cross the network, diff it against the approved list, and enforce the allowlist at launch. The client-by-client config paths are in how to find every MCP server on your laptops, and the Agent Exposure Report inventories a pilot cohort first. |
| MCP10Context Injection & Over-Sharing | Shared, persistent, or insufficiently scoped context windows let what one task, user, or agent saw leak into another session. | Scope context per division and role with one ACL enforced on both reads and writes, run content-safety on ingest and recall, and scan outbound payloads for data that should not leave the session. Constellation is the shared-memory plane doing this today. |