Agentic Attacks Begin to Remove Humans From Parts of the Kill Chain

· Original article ↗

Summary

Intrinsec cites reports of agents automating parts of cyberattacks, including a reported campaign against 27 firms, Android malware PROMPTSPY, LLM-based C2 malware CLOSEDQUORUM, and cloud privilege escalation; it calls for AI-assisted defense and resilience.

Key points

  • Gambit Security analyzed a campaign using the Strix, Cairn, and Hermes harnesses for vulnerability research, exploitation, and orchestration; the article says 27 firms were compromised and over 600,000 card numbers harvested in five days, followed by an automated database wipe.
  • Google Threat Intelligence Group analyzed PROMPTSPY, Android malware whose GeminiAutomationAgent can interpret and interact with the interface; the article describes persistence using accessibility functions and the Recent Apps menu.
  • Cisco Talos detected CLOSEDQUORUM, a Go malware that uses a plurality vote among DeepSeek, Gemini, and Mistral models to select actions. Its calls to legitimate AI-provider APIs can make domain blocking ineffective, according to the article.
  • A Microsoft report describes Storm-3168, also known as JADEPUFFER, using agents to navigate cloud environments and escalate privileges through compromised service principals.
  • Intrinsec argues that agentic attacks are not yet fully autonomous end to end, but could outpace conventional response cycles; it advocates AI-assisted defense and greater business resilience.

Article Details

Event Type
Reported use of autonomous AI agents in cyberattacks, including reconnaissance, exploitation, persistence, command and control, and cloud privilege escalation.
Impact
According to Gambit Security's analysis, an unnamed cybercrime actor compromised 27 online retailers in five days, collected more than 600,000 payment card numbers, and then automated a database wipe to remove traces. Separate reported examples describe PROMPTSPY using an agent to interact with an Android interface, CLOSEDQUORUM using AI models for C2 decision-making, and Storm-3168 using agents to navigate cloud environments and escalate privileges. The article warns that the speed and persistence of agent-assisted attacks can challenge conventional defenses.

MITRE ATT&CK

Threat Actors

Malware

Products

Tools

Related Articles