Responsible researchers need a clear path to report prompt injection, model-data exposure and tool misuse without guessing whether their test will trigger an account ban or legal threat.
Publish scope and safe harbor carefully
Specify covered domains and AI endpoints, excluded third-party infrastructure, acceptable testing intensity and prohibited handling of real personal data. Provide a security contact, encrypted submission option and expected initial acknowledgment. Safe harbor language must be reviewed by counsel and cannot bind unrelated owners.
Include AI-specific evidence
Ask for affected model or feature version, content source, tool or retrieval path, authorization context, steps to reproduce and impact. Do not request a complete stolen dataset merely to prove leakage. Accept redacted examples and coordinate secure transfers where truly needed.
Triage by real-world impact
A prompt that makes the system say something odd is not necessarily a vulnerability. Prioritize unauthorized data access, uncontrolled tool actions, persistent poisoning, privacy compromise and reproducible cross-user effects. Track exploit preconditions and mitigate root permissions rather than only filtering one known phrasing.
Close the feedback loop
Acknowledge findings, identify duplication safely, schedule validation, communicate when a fix is deployed and give attribution with the researcher’s consent. Retest after model, tool and retrieval upgrades. CISA’s 2020 federal directive demonstrates disclosure program design but applies to the specified US agencies, not the entire AI industry.
Example
An independent researcher finds that a public chatbot reveals a private CRM field through an unauthorized lookup. The program uses a synthetic record for proof, disables the unsafe tool privilege, informs affected owners and publishes a sanitized advisory after correction.
Sources & Further Reading
Evidence checked 25 October 2026. Applicability: Global; US federal VDP example. This material is general education, not legal, medical, investment or other professional advice. No vendor performance test is implied.