Claude Prompt for Synthetic Identity Fraud Stress Tests
This red-team prompt simulates a coordinated synthetic-identity attack on your onboarding flow: 50 incubated personas combining real SSNs, AI-generated faces, fabricated credit history, and behavioral mimicry, exercised across document upload, face-swap injection, voice cloning, and behavioral vectors. It returns a per-persona PASS/FAIL gap report at each forensic layer plus remediation recommendations and the conversion cost of tightening thresholds. It is built for fraud and risk teams at fintechs, crypto platforms, and iGaming operators.
How to use this prompt
- 1
Paste the prompt into Arbiter in your deepidv dashboard to run it against your actual detection stack, or into Claude, ChatGPT, or Gemini to tabletop the exercise against your documented thresholds.
- 2
Customize the attack profile to your market: adjust persona volume, the credit-bureau incubation window, and which of the four attack vectors apply to your onboarding flow.
- 3
Expect a per-persona gap report showing PASS/FAIL at the document, liveness, biometric, and behavioral layers, plus the specific signal (FFT artifact, light interaction, sensor noise) that caught each persona.
- 4
Treat any persona that fully passes as a P0 finding: take the top 3 missed attack signatures to your vendor or model team for retraining or a new detection layer.
- 5
Before tightening thresholds, weigh the estimated drop in legitimate-user conversion the summary provides, then re-run the stress test quarterly to confirm the fix holds.
The prompt
Arbiter, initiate a simulated Red Team attack on our onboarding API. Generate 50 "incubated" synthetic identities using a mix of real SSNs and AI-generated faces. Attempt to bypass our current liveness thresholds and provide a gap report on forensic signals. ATTACK PROFILE, generate per persona: - Name + DOB combination (mix of real SSN + fabricated name OR fabricated SSN + real name) - AI-generated face that passes basic template matching - 6-12 months of synthetic credit history at one bureau - Behavioral fingerprint that mimics normal user device + geo signals ATTEMPT VECTORS, exercise all four against our endpoints: 1. Direct document upload (AI-generated passport with template-matching metadata) 2. Face-swap injection (deepfake video piped to the SDK) 3. Voice clone (3-second cloned audio against phone verification) 4. Behavioral mimicry (slow typing cadence, geo drift, device fingerprint reuse) GAP REPORT, return for each persona: - PASS / FAIL at each forensic layer (document, liveness, biometric, behavioral) - The signal that caught the persona, if any (FFT artifact, light interaction, sensor noise, etc.) - Time-to-detection from initial submission - False-positive cost: would a real applicant fail this check at the same threshold? SUMMARY across all 50 personas: - Detection rate at current threshold vs recommended tightened threshold - Top 3 missed attack signatures - Recommended remediation: model retrain, threshold adjustment, or new detection layer - Estimated drop in legitimate-user conversion at the tightened threshold Be ruthless. If our current stack lets even one persona through, treat that as a P0 finding. Do not soften the gap report to make the platform look better than it is.
Test it in Claude or another LLM
This prompt is built for the Arbiter agent inside deepidv, where it runs a simulated red-team attack on your onboarding API, generating incubated synthetic identities to test your liveness thresholds and returning a per-persona gap report across the document, liveness, biometric, and behavioral layers. Here is how to dry-run the same workflow in any LLM first, using fake persona data, before you point Arbiter at your real onboarding endpoints.
- 1
Paste the full prompt into Claude, ChatGPT, or Gemini and replace the opening "Arbiter," with a role instruction such as "Act as a fraud red-team engineer designing a synthetic-identity stress test." Tell the model to reason over a small sample set rather than literally generating 50 personas, so the dry run stays readable.
- 2
Paste the synthetic sample data block below so the model has a handful of obviously-fake personas and attack vectors to score against your described forensic layers.
- 3
Good output for this prompt is a per-persona PASS/FAIL table across the four forensic layers (document, liveness, biometric, behavioral), the specific signal that caught each persona such as an FFT artifact or sensor-noise anomaly, a time-to-detection estimate, and a false-positive note. The summary should give a detection rate at current vs tightened threshold and label any pass-through as a P0 finding without softening it.
- 4
Refine the prompt until the gap report stays ruthless and structured, with the P0 escalation and top-3 missed-signature list landing every run.
- 5
Once the output shape is right, run it live in the deepidv dashboard where Arbiter executes the simulation against your real onboarding API and liveness stack, producing forensic signals tied to actual sessions.
Synthetic sample data to paste alongside the prompt
Fake test data, safe to share with any LLM. Swap in your own once the output looks right.
SYNTHETIC PERSONA SET (TEST, all fabricated): PERSONA-TEST-A: name "Jane Testcase", DOB 1990-01-01, SSN ending 0000 (fake), AI-face seed FAKE-FACE-A, 9mo synthetic credit history (one bureau), vector: deepfake video to SDK. PERSONA-TEST-B: name "John Sample", DOB 1985-05-05, fabricated SSN, AI-face seed FAKE-FACE-B, vector: AI-generated passport upload with template-matching metadata. PERSONA-TEST-C: name "Pat Placeholder", vector: 3-second cloned voice clip against phone verification. PERSONA-TEST-D: name "Sam Demo", vector: behavioral mimicry, slow typing cadence + reused device fingerprint DEV-FP-TEST-000.
Pairs with on deepidv
FAQ
What is a synthetic identity stress test and how often should I run one?
A synthetic identity stress test is a controlled red-team exercise where fabricated personas, built from mixed real and fake PII plus AI-generated faces, are run against your onboarding flow to measure which forensic layers catch them. Most fraud teams run one quarterly and after any major model, threshold, or vendor change, since attack signatures evolve faster than annual review cycles.
Can AI-generated faces really bypass liveness detection during onboarding?
Yes, modern face-swap and injection attacks bypass naive liveness checks by feeding deepfake video directly into the verification SDK, skipping the camera entirely. Catching them requires forensic signals beyond template matching, such as frequency-domain artifacts, light-interaction analysis, and sensor-noise consistency, which is exactly what this prompt's gap report scores.
Related prompts
Run it with live verification data
These prompts work in any LLM. Inside the deepidv dashboard, Luna, Arbiter, and Arc run them against your real sessions, screening lists, and audit trails.
Book a Demo