A compromised fraud detection platform lets an attacker or a dishonest insider decide which payments, logins and applications are approved, while the platform's own dashboards keep reporting that fraud controls are working. This playbook covers the response when the platform's console, rules, lists, models or API credentials can no longer be trusted: how tampering shows up, the first hour of triage, containment that does not leave customers unprotected, credential rotation, rebuilding rules and models from a known-good baseline, and regulatory reporting. It applies to vendor services and self-hosted platforms used for payments, account takeover, onboarding and payouts. It is written for fraud operations leads, platform owners and incident responders, and it is vendor-neutral. Preventive controls are in Securing the Fraud Detection Platform Itself, and this page does not repeat them.
Key takeaways
- Integrity is the main loss, not just confidentiality. The question is which decisions were wrong and since when, so reconstruct the rule, list and model state for the whole window.
- Do not fail open by accident. Switching the platform off may approve everything; contain with a frozen known-good rule set or conservative fallback rules agreed per channel.
- Treat insiders as a live possibility. Self-dealing allowlist entries and closed cases are a common pattern, so keep the investigation team separate from the people whose changes are under review.
- Rotate every machine credential the platform touches. Decision API keys, ingestion accounts, outbound action credentials and webhook secrets all need new values, one per service.
- Corrupted labels poison future models. Outcome labels changed during the window must be corrected before the next model is trained or promoted.
What an attacker can change, and what it costs
| Component reached | Abuse | Business effect | First response |
|---|---|---|---|
| Rules and thresholds | Disable rules, raise score thresholds, add narrow exceptions | Chosen fraud is approved | Freeze rule changes, restore last approved version |
| Allowlists and blocklists | Allow attacker cards, devices, accounts or merchants; block legitimate customers | Targeted approval or denial of service | Export lists, remove entries without approval |
| Models and feature store | Promote a weakened model, poison training data | Gradual rise in missed fraud | Roll back to the previous model artefact |
| Case management and labels | Close alerts, relabel fraud as genuine, export customer data | Lost recoveries, corrupted training, data breach | Lock bulk actions and exports, preserve case history |
| Decision API credentials | Probe thresholds, submit false outcomes | Attackers learn what passes | Revoke and reissue per calling service |
| Outbound action credentials | Mass card blocks, account locks or false notifications | Customer denial of service, social engineering | Disable the action integration, rotate credentials |
| Data feeds and consortium links | Inject false events or exfiltrate shared data | Wrong decisions, data protection breach | Suspend feeds, validate sources |
Relevant techniques: ATT&CK T1078 Valid Accounts and T1098 Account Manipulation for console access, T1565.001 Stored Data Manipulation for tampering with rules, lists and case outcomes, T1552 Unsecured Credentials for decision API keys found in code, and MITRE ATLAS AML.T0020 Training Data Poisoning and AML.T0043 Craft Adversarial Data at the model layer.
How the compromise shows up
- Loss without alerts: chargebacks, customer fraud claims or scheme reports rise while platform alert volumes stay flat or fall.
- Change log anomalies: rules or list entries changed outside change windows, without an approver, or by an account that does not normally author rules.
- Self-dealing indicators: a staff member's own account, card, device or address appears in an allowlist, or cases about them are closed by colleagues.
- Label drift: unusual volumes of genuine outcomes from one analyst or one feed, or labels changed long after case closure.
- API patterns: a decision API key calling from new sources, or sending sequences of near-identical requests that look like threshold probing (OWASP API4:2023 Unrestricted Resource Consumption describes missing rate limits).
- Customer impact: a spike in blocked cards, locked accounts or fraud notifications that customers did not trigger.
Triage: the first hour
- Classify the access: a console account (staff or contractor), a machine credential, a vendor-side incident or code execution on a self-hosted platform.
- Fix the window from the earliest unexplained change in the rule, list, model and admin logs.
- Identify the last known-good state: the most recent approved rule set, list snapshot and model version before the window.
- Check the fallback behaviour for each channel: what happens to decisions if you freeze or disconnect the platform.
- Decide on the investigation team. If insider involvement is possible, exclude the people whose accounts made the changes and involve HR and legal early. See the Insider Threat Programme Guide.
- Estimate exposure by channel: transactions, applications or logins decided under altered logic, and their value.
Containment by priority
- Freeze and capture. Export rules with version history, lists, model registry entries, case history and admin logs to storage the platform administrators cannot change.
- Lock administrative change. Restrict rule, list and model promotion rights to a named response group with maker-checker enforced, disable implicated accounts and revoke their sessions.
- Restore known-good logic. Reinstate the last approved rule set and model, and remove allowlist entries created in the window that lack an approver. Where unsure, run conservative fallback rules rather than switching the platform off.
- Revoke machine credentials. Reissue decision API keys one per calling service, rotate ingestion and outbound action credentials, and rotate webhook signing secrets. See the Webhook Secrets Security Guide.
- Stop data leaving. Disable bulk export and report downloads for all but the response group, and suspend consortium or partner feeds until validated.
- Add compensating controls in neighbouring systems, such as step-up authentication or lower payout limits, for channels that were exposed. The PSD2/PSD3 Strong Customer Authentication Guide covers step-up options.
Rollback: keep the captured state so that any containment change that blocks too many legitimate customers can be reversed with a recorded approval, and measure false positive rates hourly while fallback rules run.
Credential rotation
| Credential | Action | Note |
|---|---|---|
| Admin and analyst accounts | Reset, enforce SSO with phishing-resistant MFA, review role assignments | NIST SP 800-53 AC-5 separation of duties |
| Decision API keys or OAuth clients | New credential per calling service; prefer short-lived tokens or mTLS where supported | Revoke one service without stopping the rest |
| Ingestion service accounts and partner API keys | Rotate, narrow scope, restrict source addresses | See the Service Account Security Guide |
| Outbound action credentials (card block, account lock, ticketing, notifications) | Rotate and limit to the specific actions; add volume alerts | Abuse causes customer denial of service |
| Model pipeline and feature store accounts | Rotate, require approval for promotion, record artefact hashes | NIST SP 800-53 CM-3 change control |
| Client-side collection keys | Rotate if the vendor supports it; treat as public regardless | Rely on server-side validation |
The API Key Management Guide and the Secrets Rotation Runbook Template cover the mechanics.
Eradication and rebuilding trust
- Rebuild rules from the approved baseline: compare the exported rule set with the last approved version line by line, and re-create only changes with a ticket and an independent approver.
- Re-verify every allowlist entry: each needs an owner, a reason, an approver and an expiry; delete the rest.
- Correct labels: list outcome labels set or changed in the window, re-review them, and exclude unverified labels from training data.
- Retrain or roll back models: do not promote any model trained on data from the window until labels are corrected; check artefact hashes against the registry. See the AI Data Poisoning and Backdoors Guide.
- Reopen closed cases closed by implicated accounts during the window and work them again.
- Self-hosted platforms with code execution: rebuild hosts from clean images and restore configuration only from reviewed exports.
Recovery and re-decisioning
- Replay the window. Run transactions and applications from the window through the restored rules and model in shadow mode to list decisions that would have differed.
- Act on the differences: review approved applications, recover or hold funds where still possible, and contact affected customers.
- Return to normal rules in stages, channel by channel, watching approval, decline and review rates against the pre-incident baseline.
- Recalculate fraud rates. Figures produced from tampered data may have been wrong, including any reported to regulators or schemes.
- Exercise the fallback again within a quarter, so the next incident does not depend on improvisation.
Evidence and notification
- Preserve: rule and list version history, model registry and artefacts, case history including label changes, admin and API access logs, and decision logs for the window. Keep chain of custody where an insider may face disciplinary or criminal action. See the Fraud Case Investigation Template.
- GDPR Article 33: where personal data was exported or exposed, the controller notifies the supervisory authority without undue delay and, where feasible, within 72 hours of becoming aware, unless the breach is unlikely to result in a risk to individuals.
- PSD2 transaction risk analysis: under Commission Delegated Regulation (EU) 2018/389, payment service providers must run transaction monitoring (Article 2) and, if they use the transaction risk analysis exemption, must immediately report to competent authorities when a monitored fraud rate exceeds the reference rate (Article 20). The exemption is suspended after two consecutive quarters above the reference rate. Fraud rates calculated from a tampered platform need checking for this reason.
- DORA: EU financial entities report major ICT-related incidents to their competent authority with an initial notification, intermediate report and final report under Article 19 of Regulation (EU) 2022/2554. See the DORA Executive Briefing.
- Other regimes: see the Cyber Incident Reporting Requirements Map.
Decision guidance
| Situation | Choose | Trade-off |
|---|---|---|
| Single insider changing allowlists | Revoke access, remove entries, reopen cases, preserve evidence for HR and legal | Platform keeps running; needs a clean investigation team |
| Machine credential stolen, no console access | Rotate the credential, review API logs for probing or false outcomes | Short disruption to the calling service |
| Console admin compromise with rule and model changes | Freeze, restore baseline, rotate everything, replay the window | Higher false positives while fallback rules run |
| Vendor-side incident | Rotate your credentials, verify your own rules and lists, follow the vendor's scope statement | Dependence on the provider's investigation |
| Small team without a fallback rule set | Restore the last export and add step-up authentication or manual review for high-value transactions | Manual workload and customer friction for a period |
Detection and measurement
- Time to known-good: hours from declaration to restored, approved rules and model in production.
- Change approval coverage: rule, list and model changes with an independent approver; target 100%.
- Loss to alert ratio: confirmed fraud losses compared with alert volumes per channel, which exposes silent suppression.
- Credential sharing: calling services sharing a decision API credential; target zero.
- Re-decision completion: share of window decisions replayed and reviewed.
Forward rule, list, model, label and admin events to a SIEM the platform administrators cannot alter (NIST SP 800-53 AU-9). See the Fraud Metrics and KPIs Guide for operational fraud measures.
Framework mapping
| Framework | Reference | Where it applies |
|---|---|---|
| NIST CSF 2.0 | RS.MA-01, RS.AN-03, RS.MI-01, RS.MI-02, RS.CO-02, RC.RP-05 | Coordinated response, analysis, containment, eradication, notification, verified restore |
| NIST SP 800-53 Rev. 5 | IR-4, IR-6, AC-5, IA-5, CM-3, SI-7, AU-9 | Incident handling and reporting, separation of duties, credentials, change control, integrity, protected logs |
| EU payments | Delegated Regulation (EU) 2018/389 Articles 2, 20, 21 | Transaction monitoring, fraud rate reporting and exemption suspension |
| MITRE ATT&CK | T1078, T1098, T1565.001, T1552 | Console access, role changes, data tampering, exposed keys |
| MITRE ATLAS | AML.T0020, AML.T0043 | Training data poisoning, adversarial inputs |
| OWASP API Security Top 10 2023 | API2, API4 | Broken authentication, unrestricted resource consumption on decision APIs |
Common failure patterns
- Switching the platform off. Channels fail open and the incident becomes a fraud loss event.
- Investigating with the suspects. People whose changes are under review keep access to the case data and logs.
- Rotating one shared key. Every channel used the same decision API key, so rotation breaks all of them at once and nobody can tell which caller was abused.
- Retraining on poisoned labels. The next model learns the attacker's relabelled outcomes and the problem returns.
- Forgetting reported figures. Fraud rates already sent to regulators or schemes were calculated from tampered data and are not corrected.
- No replay. Wrongly approved applications and accounts stay open because nobody re-decided the window.
Practitioner checklist
- Keep regular exports of approved rules, lists and model versions outside the platform.
- Agree and test fallback behaviour per channel before an incident.
- On compromise, capture state, then lock change rights to a response group with maker-checker.
- Restore known-good rules and models; remove unapproved allowlist entries.
- Reissue decision API keys per calling service and rotate all integration and outbound credentials.
- Disable bulk export and suspend partner feeds until validated.
- Correct labels from the window before any retraining.
- Replay window decisions in shadow mode and act on the differences.
- Check GDPR, DORA and PSD2 fraud rate reporting duties.
- Keep implicated staff out of the investigation and preserve chain of custody.
Standards and references
- NIST: SP 800-61 Rev. 3, Incident Response Recommendations and Considerations for Cybersecurity Risk Management
- EUR-Lex: Commission Delegated Regulation (EU) 2018/389, RTS on strong customer authentication
- EUR-Lex: Regulation (EU) 2022/2554, Digital Operational Resilience Act
- EUR-Lex: Regulation (EU) 2016/679, General Data Protection Regulation
- MITRE ATT&CK: Stored Data Manipulation
- MITRE ATLAS: Training Data Poisoning
- OWASP: API Security Top 10 2023
Related NHI Mgmt Group resources: Fraud Prevention Guide for Digital Businesses · Fraud Operations and Case Management Guide · If your payment gateway is compromised · If your identity verification and KYC platform is compromised