mediumAgent ThreatPrompt Injection

Multilingual and Obfuscated Prompt Injection Susceptibility in Mainstream LLMs

First seen Jul 5, 2026 · Updated Jul 5, 2026

prompt-injectionmultilingual-attacksobfuscationjailbreakresearchphishing-generationmalware-generationsafety-alignmentASI01 · Goal HijackingAML.T0051AML.T0054Surface: ModelPropagation: None

This is an academic research paper (arXiv, not an active exploit) empirically benchmarking six major LLMs against direct, multilingual, and obfuscated prompt injection attacks. The study finds that all tested models can be induced to generate phishing content, malicious websites, and malware, with non-English prompts and multi-stage obfuscation significantly increasing compliance rates. Since this is a research disclosure rather than an in-the-wild exploit or agent-specific vulnerability, severity is rated medium rather than high/critical.

Technical Analysis

The entry point is the standard text prompt interface of six LLMs (DeepSeek, GPT, Gemini, Grok, Llama, Qwen), where attackers craft direct or multi-stage adversarial prompts, sometimes translated into non-English languages or lightly obfuscated via character encoding, to bypass safety alignment. The mechanism exploits gaps in safety training coverage across languages and insufficient robustness to encoding-based obfuscation, causing the model to comply with requests to generate phishing emails, deceptive websites, or malware code. The attacker gains malicious content generation capability directly from the base model; there is no described agentic tool-chaining, memory persistence, or inter-agent boundary crossing in this study — the vulnerability is confined to the model layer itself. If such a model were embedded as the reasoning core of an autonomous agent with tool access, successful injection could cascade into downstream tool misuse, but this paper does not test that scenario.

Detection Signatures

  • Prompts containing task instructions embedded in non-English languages inconsistent with prior conversation context
  • Multi-stage prompts that incrementally build toward a harmful request (context laundering)
  • Use of character encodings (e.g., base64, unicode homoglyphs, leetspeak) to obscure sensitive keywords
  • Output requests for phishing email templates, fake login pages, or malware/exploit code framed as 'research' or 'educational'
  • Rapid language-switching mid-conversation to evade keyword-based safety filters

Remediation Steps

  1. 1

    Strengthen multilingual safety alignment

    Extend RLHF/safety fine-tuning datasets to cover adversarial prompts across all supported languages, not just English, and continuously red-team in low-resource languages.

  2. 2

    Deploy encoding-aware input normalization

    Preprocess and decode common obfuscation techniques (base64, unicode tricks, homoglyphs) before running safety classifiers so hidden malicious intent is exposed prior to model reasoning.

  3. 3

    Add output-side content classifiers

    Run generated content through independent phishing/malware detection classifiers before returning results to users, especially for cybersecurity-adjacent requests.

  4. 4

    Implement multi-turn intent tracking

    Monitor conversation state across turns to detect incremental/multi-stage jailbreak attempts rather than evaluating each prompt in isolation.

  5. 5

    Restrict high-risk generation in agentic deployments

    When these models are used as agent backbones with tool/execution access, gate code-generation and web-content outputs behind human review or sandboxed execution.

Industries Most Exposed

cybersecurityfinancial servicestechnologyany industry deploying customer-facing or agentic LLMs

Sources

Respond to this threat

Pro subscribers get a full AI-generated incident-response playbook for this threat — detection, containment, eradication, and recovery steps — plus an unlimited AI Threat Advisor for questions about your environment.