highAgent ThreatPrompt Injection

LoginTrap: Phishing-Style Indirect Prompt Injection Inducing Web Agent Logins

First seen Aug 6, 2026 · Updated Aug 6, 2026

web-agentindirect-prompt-injectionphishingcredential-theftauthentication-bypassfuzzingblack-box-attackbrowser-automationASI01 · Goal HijackingAML.T0051AML.T0054Surface: Human InterfacePropagation: None

Researchers demonstrate LoginTrap, an indirect prompt injection technique that manipulates LLM-based web agents into believing login is a necessary step to complete a task, redirecting them to attacker-controlled login pages. This exploits the authentication boundary of web agents to potentially exfiltrate user credentials, achieving an 86% average success rate across multiple LLM backbones and agent architectures. This is a research disclosure, not an observed active exploit, but it demonstrates a credible and highly effective attack surface.

Technical Analysis

The attack embeds page-specific indirect prompt injection content into webpages that web agents visit while performing autonomous tasks; a fuzzing-inspired generation process crafts context that convincingly frames login as a plausible prerequisite for task continuation. The attacker operates in a black-box model, controlling only webpage content and the induced login flow, without needing knowledge of the user's task or the underlying agent architecture/model. Because web agents interpret webpage content as part of their operating context (mixing untrusted data with instructions), the injected content hijacks the agent's plan and steers it toward attacker-controlled login pages, crossing the boundary between passive content observation and privileged credential-entry action. The attacker's gain is potential capture of user credentials or sensitive authentication data, effectively turning the agent into an unwitting proxy for credential exfiltration; the technique remains effective across different LLM backbones, agent architectures, and existing defenses.

Affected Systems

generic LLM-based web agent frameworks (browser-automation agents)

Detection Signatures

  • Webpage content containing unsolicited login prompts unrelated to the user's original task intent
  • DOM elements or injected text framing login as a 'required step' or 'prerequisite' inserted mid-task
  • Agent navigation logs showing unexpected redirection to login/credential-entry pages not part of original task plan
  • High-frequency A/B variations of login-inducing phrasing across pages (fuzzing artifacts)
  • Agent action traces where login attempts occur on domains not matching the original task's target site

Remediation Steps

  1. 1

    Authentication-aware action gating

    Require explicit user confirmation or out-of-band approval before a web agent submits credentials or navigates to any login page, especially domains differing from the original task target.

  2. 2

    Content-instruction separation

    Architect agents to treat webpage text as untrusted data, not as instructions; use structured extraction and instruction-filtering layers to strip imperative or task-redirecting language from observed page content.

  3. 3

    Domain allow-listing for credential entry

    Restrict agents to only perform login/credential-submission actions on pre-approved, user-verified domains associated with the original task.

  4. 4

    Task-goal consistency checks

    Implement a planner-level check that flags and halts execution when a proposed action (e.g., login) is not a coherent, previously anticipated step of the user's original task.

  5. 5

    Red-team with fuzzing-style injection testing

    Proactively test agents against fuzzing-generated indirect injection payloads (like LoginTrap) across diverse page contexts and agent backbones before deployment.

Industries Most Exposed

e-commercefinanceSaaSidentity-managementgeneral web services

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.