Skip to content

Chronicle YARA-L: AI Agent Indirect Context Injection & Markdown Exfiltration — Threat Hunting Query

Density:
Evaluation DimensionObserved Telemetry / AttributeEpistemic Status
Hunting Query IdentifierHUNT-CHRONICLE-LLM-PROMPT-INJECTIONCanonical Hunt Node
Target SIEM / Analytics PlatformGoogle Cloud Chronicle SecOpsTelemetry Ingestion Engine
Platform Syntax Dialectchronicle_yaralQuery Language Standard
Telemetry & Log SourcesApplication Gateway Telemetry, Agent Execution Traces, DNS Query LogsData Sensor Sources
Targeted Adversary SyndicatesCLOSEDQUORUM Threat ClusterThreat Attribution
Associated Weaponized CVEsCVE-2026-101861Vulnerabilities Targeted
Associated Intrusion CampaignsCAMP-GLOVE-2026Monitored Campaign
Analytical Confidence95%Hypothesis Validity Score
Primary Evidence SourcesSRC-CISCO-TALOS, SRC-GITHUB-ADVISORIESIntelligence Attestation

Adversaries inject instructions into retrieved RAG documents that command autonomous agents to embed sensitive API keys inside markdown images fetched from attacker servers.

Adversary operations frequently slip through standard alerting thresholds by leveraging legitimate binaries or abusing valid service configurations. Proactive threat hunting formulates precise hypotheses based on known architectural chokepoints and queries historical telemetry to unearth stealthy intrusion footholds.


rule hunt_llm_markdown_image_data_exfiltration {
meta:
description = "Hunts for outbound image rendering requests carrying base64 payloads initiated by AI agent runtime services"
author = "Hermes Detection Engineering"
severity = "HIGH"
events:
$http.target.url = /\!\s*\[.*?\]\(https?:\/\/[^\/]+\/[a-zA-Z0-9_-]+\.(?:png|jpg|gif)\?[a-zA-Z0-9_]+=([a-zA-Z0-9+\/=]{32,})\)/
$http.principal.process.command_line = /langflow|llama-index|langchain|autogen/
condition:
$http
}

Isolate the agent orchestration runner, review RAG ingestion history for recently indexed external web pages, and decode the base64 GET query parameters to identify stolen secrets.

  1. Verify Timeline Baseline: Review the first seen and last seen timestamps of returned events to distinguish chronic benign traffic from sudden operational spikes.
  2. Examine Parent-Child Lineage: Ensure that all parent processes match authorized deployment manifests and do not conceal unexpected interactive utilities.
  3. Inspect Network Pivots: Cross-reference communicating source and destination IP addresses against Hermes Adversary Infrastructure endpoints.