{
  "id": "talos-a3-ai-analysis-evasion-2026-10",
  "name": "A3 AI-analysis evasion (FRUITSHELL / HOLLOWCLAD / MANTLEMAZE)",
  "date_reported": "2026-10-08",
  "researcher": "Ryan Fetterman",
  "org": "Cisco Talos (CAIRN)",
  "sources": [
    {
      "url": "https://blog.talosintelligence.com/ignore-all-instructions-and-read-this-blog-the-state-of-ai-analysis-evasion-in-malware/"
    }
  ],
  "magicsword_coverage": [],
  "borderline": false,
  "borderline_note": null,
  "ai_role": "other",
  "ai_role_secondary": null,
  "ai_role_detail": "Malware embeds natural-language instructions to steer defender LLM triage (prompt injection against analyzers)",
  "technique_summary": "84 samples across families embed plaintext comments telling LLMs to ignore the file, refuse reverse engineering, or claim DoD/IP. Direct-instruction trick steered models toward benign ~35% in best runs. MANTLEMAZE pairs this with iqvw64e.sys BYOVD (CVE-2015-2291); ROZESHELL pairs with AMSI bypass and csc.exe.",
  "initial_access": "unknown (varies by family; technique is anti-analysis layer)",
  "ai_tooling_and_models": [
    "Prompt injection against defender LLM pipelines"
  ],
  "other_tooling": [
    "iqvw64e.sys (LOLDrivers)",
    "csc.exe (LOLBAS)",
    "VMProtect costume sections"
  ],
  "victims_and_sectors": "unknown",
  "regions": [
    "unknown"
  ],
  "attribution_or_cluster": "CAIRN families FRUITSHELL, PLOTSAFE, HOLLOWCLAD, MANTLEMAZE, ROZESHELL; A3 technique class",
  "scale_and_metrics": "84 samples Jan 2025–Jul 2026; ~35% steering in best direct-instruction runs",
  "prevention_angle": "Treat sample text as evidence never instruction; prevent iqvw64e load via app control",
  "related_magicsword_projects": [
    "LOLDrivers",
    "LOLBAS"
  ],
  "source_notes": null,
  "last_reviewed": "2026-10-20",
  "attack_flow": {
    "nodes": [
      {
        "id": "step-1",
        "title": "Malware is presented for analysis",
        "description": "Samples include comments aimed at automated analysis models.",
        "kind": "access",
        "source_urls": [
          "https://blog.talosintelligence.com/ignore-all-instructions-and-read-this-blog-the-state-of-ai-analysis-evasion-in-malware/"
        ]
      },
      {
        "id": "step-2",
        "title": "Instructions target the analyst model",
        "description": "The text asks LLMs to ignore the file, refuse analysis or accept benign explanations.",
        "kind": "ai",
        "source_urls": [
          "https://blog.talosintelligence.com/ignore-all-instructions-and-read-this-blog-the-state-of-ai-analysis-evasion-in-malware/"
        ]
      },
      {
        "id": "step-3",
        "title": "Model interpretation is influenced",
        "description": "The research measures whether direct instructions shift maliciousness assessments.",
        "kind": "execution",
        "source_urls": [
          "https://blog.talosintelligence.com/ignore-all-instructions-and-read-this-blog-the-state-of-ai-analysis-evasion-in-malware/"
        ]
      },
      {
        "id": "step-4",
        "title": "Some runs misclassify samples",
        "description": "Results are bounded by the evaluated models, samples and experimental conditions.",
        "kind": "impact",
        "source_urls": [
          "https://blog.talosintelligence.com/ignore-all-instructions-and-read-this-blog-the-state-of-ai-analysis-evasion-in-malware/"
        ]
      }
    ],
    "edges": [
      {
        "source": "step-1",
        "target": "step-2"
      },
      {
        "source": "step-2",
        "target": "step-3"
      },
      {
        "source": "step-3",
        "target": "step-4"
      }
    ]
  },
  "atlas_techniques": [
    {
      "id": "AML.T0051.001",
      "rationale": "Comments in analyzed files inject instructions into a security-analysis model.",
      "source_urls": [
        "https://blog.talosintelligence.com/ignore-all-instructions-and-read-this-blog-the-state-of-ai-analysis-evasion-in-malware/"
      ]
    },
    {
      "id": "AML.T0015",
      "rationale": "The text attempts to evade AI-based maliciousness assessment.",
      "source_urls": [
        "https://blog.talosintelligence.com/ignore-all-instructions-and-read-this-blog-the-state-of-ai-analysis-evasion-in-malware/"
      ]
    }
  ],
  "category": "adversary-use-of-ai",
  "attribution": {
    "actor_name": "CAIRN families FRUITSHELL, PLOTSAFE, HOLLOWCLAD, MANTLEMAZE, ROZESHELL"
  },
  "ai_involvement_confidence": "confirmed",
  "detections": [
    {
      "name": "Review attack-specific indicators",
      "format": "guidance",
      "notes": "Monitor for the techniques described in A3 AI-analysis evasion (FRUITSHELL / HOLLOWCLAD / MANTLEMAZE). Refer to sources for specific IOCs and TTPs."
    }
  ]
}