Hidden Instructions Buried in Client Documents

E-discovery and due diligence agents retrieve emails, contracts, and PDFs from data rooms supplied by clients and counterparties. Hidden instructions in those documents can redirect an agent's next action, skew a review, or exfiltrate privileged material. PointGuard AI scans retrieved content and tool outputs for injection patterns before they reach the model and constrains what agents can do afterward.

Firms routinely process documents written by adversaries.

• Counterparty documents can carry hidden instructions.
• Injected text can bias review or summarization.
• Agents with tool access can be told to share data.
• Manipulation is hard to detect in large document sets.

Runtime Guardrails and AI Red Teaming block injected instructions and test review agents.

1. Inspect retrieved content. Scan documents, emails, and tool outputs for injection patterns using fine-tuned detection models at sub-100 ms latency.

2. Constrain agent actions. Limit which tools review agents can call after reading untrusted content.

3. Flag suspicious documents. Route documents with detected manipulation to reviewers.

4. Test continuously. Probe agents with injection scenarios before matters begin and after updates.

• OWASP Top 10 for LLMs: LLM01 Prompt Injection, LLM02 Sensitive Information Disclosure, and LLM08 Vector and Embedding Weaknesses, covering hidden instructions, privileged data theft, and manipulated document retrieval.
• OWASP Top 10 for Agentic Applications: ASI01 Agent Goal Hijack and ASI02 Tool Misuse and Exploitation, addressing hijacked review agents and tools triggered by injected content.
• NIST AI RMF: MAP identifies untrusted inputs; MEASURE tests resistance; MANAGE blocks attacks; GOVERN assigns matter ownership.