Blocking Prompt Injection in FOIA Documents

FOIA and records agents process requests and documents submitted by the public. Attackers can hide instructions in those filings that try to override redaction rules, retrieve unreleased case files, or change how a request is handled. PointGuard AI inspects retrieved documents and tool outputs before they reach the model and enforces output policies that protect exempt information.

Records agents read content that anyone can write.

• Public filings can carry hidden or encoded instructions.
• Injected text can try to bypass redaction rules.
• Agents with repository access can retrieve unreleased files.
• Errors in released records are difficult to undo.

Runtime Guardrails and AI Red Teaming block injected instructions and test records agents before release.

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

2. Enforce redaction policies. Check outputs for exempt and sensitive information before release, and block responses that violate policy.

3. Constrain access. Limit records agents to the repositories and files relevant to each request.

4. Test continuously. Probe agents with injection and redaction-bypass scenarios before release and after updates.

• OWASP Top 10 for LLMs: LLM01 Prompt Injection, LLM02 Sensitive Information Disclosure, and LLM05 Improper Output Handling, covering hidden instructions, exposed records, and unsafe release of model output.
• OWASP Top 10 for Agentic Applications: ASI01 Agent Goal Hijack and ASI02 Tool Misuse and Exploitation, addressing hijacked agent goals and retrieval tools turned against the agency.
• NIST AI RMF: MAP identifies untrusted inputs; MEASURE tests resistance; MANAGE blocks attacks; GOVERN assigns records ownership.