/

CVE-2026-91988

Severity:

HIGH RISK

(8.1)

Cleartext HTTP (CWE-319) in Atomic Agents MCP Registry

CVE-2026-91988

Deferred

Vulnerability type

Category:

Cleartext Transmission of Sensitive Information

CWE:

CWE-319

Impact:

Remote Code Execution

Severity:

HIGH RISK

(8.1)

CVSS version:

3.1

Catch this CVE while code is being written

Detect vulnerable code directly in pull requests, so CVEs are fixed before merge.

Summary

A high-severity transport security weakness in atomic-agents-stack allows the HTTP MCP server-registry backend to accept cleartext HTTP endpoints. A network attacker positioned to intercept or alter registry traffic may rewrite catalog responses, causing the agent to spawn attacker-controlled commands and arguments as local subprocesses.

Affected context

The issue occurs in deployments using the HTTP-based MCP server-registry backend of atomic-agents-stack before secure transport requirements are enforced. Risk is greatest when registry requests traverse an untrusted or shared network and the returned catalog is consumed by MCPClientPool without strong authenticity and integrity verification.

Exploit preconditions

  • Attack surface: HTTP communication between the agent and the MCP server registry
  • Required access: Network positioning sufficient to observe and modify registry responses
  • Trigger point: Retrieval and processing of an MCP server catalog
  • Why exploitation is possible: Cleartext responses can be rewritten, and attacker-controlled command and argument values may subsequently be passed to local subprocess creation

Why this happens

The underlying bug is the acceptance of cleartext HTTP schemes for a security-sensitive service-discovery channel. Because the registry response is not protected against network tampering, an intermediary can alter server definitions before they are consumed by MCPClientPool. This maps to CWE-319: Cleartext Transmission of Sensitive Information, with the broader security failure involving insufficient authenticity and integrity protection for trusted configuration data.

Potential impact

  • An attacker could execute arbitrary commands on the agent host
  • The agent process or its runtime environment may be compromised
  • Credentials, tokens, or locally accessible data could be exposed
  • Malicious MCP server definitions could persist in operational workflows
  • Compromised agents could be used for lateral movement or supply-chain abuse

General mitigation guidance

Mitigation generally focuses on preventing insecure transport and reducing trust in remotely supplied process-launch parameters.

Common defensive patterns include:

  • Restricting registry communication to authenticated, integrity-protected transport
  • Rejecting cleartext HTTP endpoints by default
  • Validating registry responses against an explicit schema and trusted policy
  • Allowing only approved executables and constrained argument formats
  • Separating catalog discovery from process execution authorization
  • Running agent processes with least privilege and limited filesystem and network access
  • Using network controls that prevent downgrade or interception of registry traffic

How teams detect issues like this

Organizations can look for this weakness through:

  • Static analysis rules that identify HTTP URL acceptance in registry or service-discovery code
  • Code review checks for transport-scheme validation and missing server authentication
  • Tests that verify cleartext endpoints are rejected under default policy
  • Runtime telemetry for registry responses that introduce unexpected commands, arguments, or executable paths
  • Egress and network policies that flag or block cleartext requests from agent workloads
  • Automated security review tools that inspect subprocess creation paths fed by remote configuration

Sources

SourceLink
github.comgithub.com/dep0we/atomic-agents-stack/security/advisories/GHSA-xhcr-cqfr-m3hv
vulncheck.comvulncheck.com/advisories/atomic-agents-stack-before-1.1.0-remote-code-execution-via-http-mcp

Frequently asked

questions

Potentially. An attacker generally needs network positioning that permits interception and modification of cleartext registry traffic; direct access to the agent host is not necessarily required.

It depends on the deployment. The key prerequisite is the ability to tamper with the registry connection or its responses, although access to the registry or surrounding network may also be needed.

Require authenticated, integrity-protected transport for registry communication and reject cleartext HTTP endpoints by default. Remote command and argument values should also undergo strict allowlisting before process execution.

Treat service catalogs as untrusted input, enforce transport security in code and policy, minimize subprocess privileges, and continuously test that network configuration cannot downgrade to cleartext communication.

Find out what's already
exploitable in your codebase.