\
Policy Center
Responsible Deepfake Testing Policy
Last Updated November 16, 2026
1. Introduction
Reality Defender tests how well identity and trust systems hold up against deepfakes and impersonation, and this policy sets the rules we follow when we do. We publish it so that customers, the companies whose products we study, and the public know what we will and will not do.
Attackers already use synthetic voices, faces, and video to get past verification checks. Understanding how those attacks succeed is central to building detection that stops them. That research has to be done carefully, lawfully, and in good faith.
We work in two ways. Authorized testing examines a specific organization’s systems with its written permission. Independent market research studies publicly available products as an ordinary user would, without attempting to defeat their safeguards. Different rules apply to each, as described below.
This policy was developed with input from Reality Defender’s independent Ethics Committee.
2. Scope
This policy covers only deepfake and impersonation risks in verification and trust workflows. That includes identity verification and onboarding, liveness checks, account security and recovery, voice and video authentication, AI agents or avatars that present a trusted identity, and content trust systems.
It does not cover general security audits, penetration testing, or broad AI red-teaming. We do not offer those services under this program.
Key terms used in this policy:
- Synthetic media: audio, images, video, or text generated or substantially altered by AI.
- Likeness: a face, voice, name, or persona that could reasonably be understood as a specific real person.
- Synthetic persona: an invented identity, with an AI-generated face or voice, that does not resemble any real person.
- Vulnerability Disclosure Program (VDP): a public program through which an organization invites good-faith security research and offers protections to researchers.
3. Guiding Principles
Seven principles apply to all of our testing and research.
- Narrow scope. We test only deepfake and impersonation scenarios in verification and trust workflows.
- Consent over surprise. Adversarial testing of a specific organization’s systems happens only with its written authorization, or within the scope of its published VDP.
- Honest identity. Without authorization, we act openly as Reality Defender. We do not use false identities or claim affiliations we do not have.
- Respect for law and terms of service. We operate within applicable law. We do not violate a product’s terms of service unless the organization that owns it has authorized the test in writing.
- Minimize harm. We use synthetic personas or Reality Defender employees wherever possible. We use a real person’s likeness only with that person’s documented consent. We never create sexual content, involve minors, harass or expose individuals, or produce election-related manipulation.
- Reproducibility without exposure. We keep detailed internal records of our methods so our work can be reproduced and audited. We do not publish the specific tools, models, or techniques we use, so this work does not become a guide for attackers.
- Responsible disclosure. Findings are private by default and shared first with the affected organization.
4. Authorized Testing
When an organization asks us to test its systems, we act only within a written authorization signed by someone with authority to grant it. The authorization defines what is in scope: which systems and environments, which methods, the testing window, any limits, and what is off-limits.
These engagements simulate how a realistic deepfake or impersonation attack could get past the organization’s verification or trust controls. Depending on what the organization approves, testing may use synthetic personas, the likeness of a consenting individual, or more adversarial techniques. Those techniques can include attempting to bypass verification controls, using test accounts, or masking that the traffic comes from Reality Defender, as a real attacker would.
We use these methods only when the organization has explicitly authorized them in writing. If the organization withdraws consent or raises a concern during testing, we stop immediately.
Results go to the organization. Reports describe what failed and how to address it. They do not include the specific tools or techniques used, unless needed to explain a finding or agreed in advance.
5. Independent Market Research
Without authorization, we study publicly available products only as an ordinary user would, to understand how deepfake and impersonation risk shows up across the market.
In this research we may:
- Use public, generally available features with no special access.
- Create synthetic faces, voices, or avatars and use them in normal, realistic ways.
- Represent Reality Defender employees, with their consent.
In this research we do not:
- Impersonate any real person, public or private, or present Reality Defender as another organization.
- Attempt to bypass safety, security, or abuse-prevention controls, including jailbreak attempts, automated brute-force testing, or scraping beyond normal use.
- Create throwaway accounts, hide our identity behind VPNs, or violate terms of service.
- Collect biometric or other sensitive data belonging to real users.
Why these limits apply. This research is meant to observe what happens through ordinary use, not to defeat a product’s protections. Trying to get around those protections would turn market research into adversarial testing of a specific system. That is a different activity, with legal and ethical obligations to the system’s owner. Those techniques therefore require the owner’s written authorization or must fall within its published VDP.
6. Responsible Disclosure
When our work reveals a meaningful weakness, we tell the affected organization privately before anything else.
- Authorized testing: findings go to the organization that authorized the test. We do not publish or describe them without its agreement.
- Independent research: if we find a serious issue and the organization runs a VDP, we report it through that program and follow its rules. If there is no VDP, we seek legal guidance on the right way to notify the organization.
- Public reporting: anything we publish from independent research is aggregated and anonymized by default. We name a specific company only after legal review, and ordinarily only after giving that company the chance to address the issue.
7. Data Handling and Provenance
We handle test data, especially biometric data such as face and voice, with care and keep it only as long as needed.
- Test media and results are stored in restricted, encrypted systems with role-based access.
- For authorized testing, retention and deletion periods are set in the written authorization.
- Biometric data is handled under applicable biometric privacy laws and destroyed once its purpose is met.
- Synthetic media we create for testing is labeled, and where feasible signed with C2PA Content Credentials, so it can be identified as synthetic.
8. Governance and Contact
This policy is owned jointly by Reality Defender’s Safety Research team and Legal, and reviewed by our independent Ethics Committee. We will update it as laws, industry standards, and our work evolve, and revise the date at the top of this page when we do.
If you believe Reality Defender has tested your product, or you have questions about this policy, contact us at security@realitydefender.com.