Skip to content

Sigma: Agentic AI MCP Semantic Tool Poisoning & Schema Override — Detection Rule

Density:
Evaluation DimensionObserved Telemetry / AttributeEpistemic Status
Detection Rule IdentifierDET-SIGMA-AI-AGENT-MCP-POISONCanonical Rule Node
Detection EngineSIGMAMulti-Format Standard
Severity ClassificationHIGHCalibrated Alert Priority
Target Technology SurfaceModel Context Protocol (MCP) Runtimes & Agent EcosystemsMonitored Platform
Test & Validation StatusVALIDATEDVerification Assurance
False-Positive Risk IndexMEDIUMOperational Noise Rating
Associated Weaponized CVEsCVE-2026-101861Target Vulnerabilities
Attributed Adversary SyndicatesCLOSEDQUORUM Threat ClusterThreat Origin
Linked Intrusion CampaignsCAMP-GLOVE-2026Active Wave
Linked Exploit ChainsAgentic Tool Execution Injection to Kubernetes Cluster TakeoverChained Topology
Analytical Confidence95%Rule Soundness Score
Primary Evidence SourcesSRC-CISCO-TALOS, SRC-GITHUB-ADVISORIESIntelligence Attestation

Monitors filesystem integrity on developer workstations and AI agent runner hosts to detect unauthorized insertion of suspicious tool executables inside MCP configuration manifests.

Detection engineering requires proactive alignment between known offensive tradecraft and defensive sensor placement. This rule intercepts adversarial attempts before lateral proliferation or host compromise can solidify.


title: MCP Tool Definition Schema Modification by Untrusted Process
id: c8945b12-9901-4188-b712-202609070001
status: stable
description: Detects modifications to Model Context Protocol configuration files or JSON-RPC tool declarations from non-IDE binaries.
references:
- https://hermes-codex.org/agentic-attack-patterns/aap-004-semantic-tool-poisoning/
author: Hermes Detection Engineering
date: 2026-09-29
logsource:
category: file_event
product: all
detection:
selection_target:
TargetFilename|endswith:
- '/mcp.json'
- '/mcp_config.json'
- '/claude_desktop_config.json'
selection_content:
EventData|contains:
- '"tools":'
- '"command": "curl"'
- '"command": "nc"'
- '"command": "powershell"'
filter_trusted:
Image|endswith:
- '/code'
- '/cursor'
- '/windsurf'
condition: selection_target and selection_content and not filter_trusted
falsepositives:
- Custom local developer script provisioning environments
level: high

Enforce digital signature verification on MCP server definitions and lock agent capability configurations into read-only filesystems.

  1. Isolation Directive: Immediately isolate hosts matching this alert from the internal VLAN to prevent lateral pivoting.
  2. Memory & Log Preservation: Capture volatile RAM dumps and forward local daemon journals to an immutable central repository.
  3. Patch Verification: Validate that all emergency vendor updates associated with CVE-2026-101861 are deployed across all edge gateways.