mediumAgent ThreatPrompt Injection

Prompt Injection Embedded in Legal Filing Text

First seen Aug 31, 2026 · Updated Aug 31, 2026

prompt-injectionlegal-techdocument-poisoningLLM-judgeindirect-injectionASI01 · Goal HijackingAML.T0051Surface: Human InterfacePropagation: Single Hop

A person reportedly embedded hidden AI instructions inside a legal filing, attempting to manipulate any AI system that later reads and processes the document into ruling or advising in their favor. This is a classic indirect prompt injection attack applied to a real-world professional document workflow rather than a novel technical exploit. Severity is moderate since the report describes a single documented incident without technical detail on payload sophistication or actual impact on a legal outcome.

Technical Analysis

The entry point is a text document (legal filing) that is ingested by an AI system—likely a legal-review, summarization, or drafting assistant used by a court, opposing counsel, or a judge's staff. Hidden instructions (e.g., disguised as invisible text, metadata, unusual formatting, or embedded in a footnote) are designed to be parsed by an LLM but overlooked by human reviewers, exploiting the gap between human and machine perception of the same document. If the downstream AI lacks separation between trusted instructions and untrusted document content, it may follow the injected directive (e.g., 'rule in favor of the filer') and produce biased summaries, recommendations, or drafted rulings, crossing the boundary from passive document analysis into influencing a human decision-maker's judgment. The attacker gains potential influence over an automated or human-assisted legal process without needing any system compromise—only that a downstream agent process the poisoned document as trusted input.

Detection Signatures

  • Presence of white-on-white or hidden text in PDFs/documents
  • Unusual instruction-like phrases within filings (e.g., 'ignore previous instructions', 'as an AI you should', 'rule in favor of')
  • Text embedded in metadata, alt-text, or non-rendering document layers
  • Anomalous formatting artifacts inconsistent with standard legal filing templates
  • LLM output that references content not visible in the rendered/printed version of a document

Remediation Steps

  1. 1

    Sanitize document ingestion

    Strip or flag hidden text, metadata, and non-visible layers before passing legal documents to any LLM-based review or summarization tool.

  2. 2

    Enforce instruction/data separation

    Ensure AI systems treat filing content strictly as data, never as executable instructions, using structured prompts that clearly delineate untrusted document text.

  3. 3

    Human-in-the-loop verification

    Require human review of AI-generated legal summaries or recommendations against the original rendered document before any action is taken.

  4. 4

    Content provenance checks

    Use document parsing tools that render text visually identical to how a human would perceive it, exposing discrepancies from hidden injected content.

  5. 5

    Audit and logging

    Log AI inputs and outputs for legal-review tools to enable post-hoc detection of anomalous instruction-following behavior.

Industries Most Exposed

legalgovernmentprofessional 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.