Recruitment technology
Prompt injection in CV screening: practical AI safeguards
A CV is evidence supplied by an applicant, not an instruction manual for the recruitment system. If a document contains text telling an AI tool to ignore criteria or change a score, that text should not control the workflow. Treat this as a document-handling risk and test it deliberately rather than assuming a polite system instruction is enough.
Separate document content from trusted instructions
OWASP’s prompt injection guidance recommends layered controls including separation of instructions and data, output validation and limited privileges. In a recruitment workflow, keep role requirements and application content in clearly defined inputs. Do not let text extracted from a CV select recipients, change permissions or trigger an outbound message. A comparison result should remain a reviewable assessment, not an authority to act.
Build an adversarial fictional test set
Create sample CVs containing a visible request to alter the outcome, a misleading claim to be a system message and instructions embedded in unusual formatting. Use the same role specification for ordinary and adversarial samples. Record what the tool produced and whether the reviewer could identify the suspicious passage. Keep the test documents fictional and separate from genuine applicant records.
Constrain outputs and downstream actions
Validate the expected report structure and reject unexpected fields or commands. Keep email approval and sending outside the model’s control. Review unusual scores against the evidence rather than trusting an apparently valid JSON object. Log enough information to investigate the event without copying unnecessary applicant information into general logs. Give staff a clear option to hold a result for technical review.
Plan for failure rather than claim immunity
No single filter should be described as complete protection. Define who can pause analysis, inspect an affected batch and retest after a change. Revisit the fictional test set when prompts, models or parsers change. Track both missed attacks and ordinary documents incorrectly flagged as suspicious; an aggressive defence that blocks legitimate applicants can create a different recruitment problem.
Sources and further reading
Sources checked 25 October 2026. The workflow recommendations are MyValidCV editorial guidance. Check current requirements with the relevant authority; this article does not establish legal compliance or certify a product.
Explore a human-reviewed CV workflow
MyValidCV Enterprise supports bulk CV-to-job comparison, evidence review and applicant email drafts for review before sending. A comparison score is advisory; your team owns hiring decisions. The operational practices described here may require separate systems and policies.
Connect your HR software
The MyValidCV API pilot brings CV-to-job comparison into your own recruitment workflow, with owner-approved company access, capped usage and human review. Request a company pilot.