Safeguarding LLMs Against Misuse and AI-Driven Malware Using Steganographic Canaries

Dimensions overall score 4.0

• The steganographic canary framework achieves 100% identifier recovery under benign and sanitization workflows (Tiers 1-2).result 95%
  Evidencepartial

  All methods achieve 100% identifier recovery under benign and sanitization workflows (Tiers 1–2).

  Implicationpartial

  Explicitly stated in the abstract with clear numeric results.

  Verificationpartial

  partial

• The hybrid Mode B (linguistic steganography augmented with symbolic layers) maintains 97% identifier recovery through targeted adversarial transforms (Tier 3).result 95%
  Evidencepartial

  The hybrid Mode B maintains 97% through targeted adversarial transforms (Tier 3)

  Implicationpartial

  Explicitly stated in the abstract with clear numeric results.

  Verificationpartial

  partial

• Improper layer composition can reduce Tier 3 identifier recovery from 97% to 0% due to cross-layer interference.technical 90%
  Evidencepartial

  We show that improper layer composition can reduce Tier 3 recovery from 97% to 0% via cross-layer interference

  Implicationpartial

  Directly stated in the analysis excerpt with specific numeric degradation.

  Verificationpartial

  partial

• The framework can detect and block canary-bearing uploads before file encryption begins in an LLM-orchestrated ransomware pipeline.result 85%
  Evidencepartial

  An end-to-end case study against an LLM-orchestrated ransomware pipeline confirms that both modes detect and block canary-bearing uploads before file encryption begins.

  Implicationpartial

  Stated in the abstract as confirmed by an end-to-end case study, though specific numeric detection rates for the case study are not provided in the excerpt.

  Verificationpartial

  partial

• Existing defenses like network DLP and endpoint controls offer limited protection once data has left an organization's control and is submitted to an LLM service.limitation 90%
  Evidencepartial

  Current defenses offer limited protection... LLM analysis of documents that have already left the organization’s control.

  Implicationpartial

  Directly stated as a limitation of current defenses in the introduction.

  Verificationpartial

  partial

• The framework supports two operational modes: Mode A marks existing documents with symbolic encodings, and Mode B generates synthetic documents using linguistic steganography.method 95%
  Evidencepartial

  We support two modes of operation where Mode A marks existing sensitive documents with layered symbolic encodings... while Mode B generates synthetic canary documents using linguistic steganography

  Implicationpartial

  Explicitly and clearly described in the abstract.

  Verificationpartial

  partial

• AI-powered malware exploits LLMs for reconnaissance, file triage, and code generation, creating a threat that converges with enterprise data leakage at the AI service ingestion boundary.market 90%
  Evidencepartial

  AI-powered malware increasingly exploits cloud-hosted generative-AI services and large language models (LLMs) as analysis engines for reconnaissance, file triage, and code generation. Simultaneously, routine enterprise uploads expose sensitive documents to third-party AI vendors. Both threats converge at the AI service ingestion boundary

  Implicationpartial

  Directly stated as the core problem motivation in the introduction and abstract.

  Verificationpartial

  partial

• This is the first framework to systematically combine symbolic and linguistic text steganography into layered canary documents for detecting unauthorized LLM processing.technical 85%
  Evidencepartial

  To our knowledge, this is the first framework to systematically combine symbolic and linguistic text steganography into layered canary documents for detecting unauthorized LLM processing

  Implicationpartial

  Explicitly claimed as a novel contribution in the abstract, though it is a knowledge claim ('to our knowledge') rather than a directly verifiable result.

  Verificationpartial

  partial

Competitive landscape

A framework using steganographic canary files to detect unauthorized LLM processing of sensitive documents and prevent AI-driven malware.