If your sanctions screening API is down, fail closed: nobody is activated, no money moves and no application is approved until a screening result for that exact person is on record. Fail open, which means treating “no answer” as “no match”, is what turns a sanctions screening API outage from a vendor incident into a violation. OFAC’s June 2026 introduction to its sanctions regime says civil penalties can rest on a strict liability standard, so in many cases a U.S. person is liable even without knowing the party was sanctioned. The maximum civil penalty it cites is $377,700 per apparent violation of the International Emergency Economic Powers Act (IEEPA), as of January 2025 and adjusted for inflation each year.
Scope: OFAC (U.S.) rules are the reference here. EU, UK and other regimes set their own duties and deadlines, so check yours. This is engineering guidance, not legal advice. Last reviewed: 25 October 2026.
What should happen if the sanctions screening API is down? Fail closed
Fail closed means the default answer to any screening failure is “not cleared”. One thing, and only one, moves a person from pending to active: a stored clear result tied to the data you actually submitted. A timeout, a 5xx error, an empty body, an unreadable payload and an unrecognised status value all produce the same outcome. That is the whole of the fail open vs fail closed screening decision, and it has to be written into code, because RFC 9110 leaves error-handling policy to the application.
“Approve and screen later” fails on legal grounds, not only on risk appetite. OFAC’s FAQ 42 states that a U.S. financial institution cannot open an account for a person on the Specially Designated Nationals (SDN) List, because that is a prohibited service, and that an opening deposit from an applicant who matches the list must be blocked. Opening first and checking afterwards means committing the prohibited act and finding out later. Every block or rejection must then be reported to OFAC within 10 business days (31 CFR 501.603 and 501.604).
Two OFAC settlements show where fail-open behaviour actually hides. It is rarely the outage itself. It is what the system does on the way out of one.
- Auto-release during a backlog. Between February 2013 and February 2018, Payoneer processed 2,220 transactions, worth $793,950.70 in total, in apparent violation of several sanctions programs. One cause, according to OFAC’s July 2021 release, was that during backlog periods flagged and pended payments were released automatically, without review. Payoneer settled for $1,385,901.40 and changed its filter to pend flagged transactions instead of letting them complete during a backlog.
- A stale watchlist after a feed failure. OFAC’s 13 October 2026 settlement with Bank of America ($11,662,930) describes two update failures. From 24 May to 28 June 2023, a vendor error left the watchlist file un-updated. On 7 February 2025, an update reflecting the previous day’s designations failed to load and was fully activated only on 12 February. OFAC called the failure to update the watchlist for over a month reckless, and stated that each organisation stays responsible for its own compliance when it relies on a third-party provider.
Sanctions API error handling: four paths, four backend decisions
Every screening call ends in one of four ways, and each needs a different response from the backend. Sanctions API error handling goes wrong when the paths get mixed up: a bad API key treated like a timeout, or a missing field treated like a hit.
| Error path | Typical signal | Account state | Retry? | Who acts |
|---|---|---|---|---|
| Technical | HTTP 400 (request malformed) or 401 (credentials refused or expired) | Pending; nothing screened | No. The same request returns the same error | Engineering, immediately |
| Timeout or outage | No response within the client timeout, a dropped connection, a 5xx | Pending | Yes: short retries, then the pending queue | Automation first; engineering alerted; compliance told if the SLA is breached |
| Validation | A required field is missing or malformed, such as no last name | Returned to the form | After the applicant corrects it | The applicant |
| Hit | A potential or exact match above your threshold | Held: no activation, no outbound funds | No. A person decides | Compliance analyst |
Validate before you call. When OFAC explains how to confirm or rule out a match, it points analysts to date of birth, nationality, identification numbers and addresses, which is exactly what a thin request lacks. Reject incomplete input at your own form, so the provider’s validation error stays an alarm rather than part of the flow.
A 401 is not one failed applicant. Every call will fail until someone fixes the credentials, so alert on it at once and never let it retry in a loop.
KYC screening timeout handling: pending queue, retry interval, user messaging
Handle a timeout in three tiers.
- In session. Retry once or twice with exponential backoff and jitter, within a few seconds. If the provider answers 429 or 503 with a Retry-After header, wait at least that long. RFC 9110 defines the header’s value as either a number of seconds or an HTTP date.
- Pending queue. When the in-session attempts fail, set the account to pending and put the application into a durable screening pending queue. A worker retries at roughly 1, 5 and 15 minutes, then every 15 minutes. These are engineering defaults, not regulatory numbers. Tune them to your provider’s measured latency and your onboarding SLA.
- Escalation. A circuit breaker stops calls after consecutive failures, so you are not hammering a provider that is down. Engineering is paged, and compliance is told when the oldest pending application passes your SLA.
Make resubmission idempotent by keying each request on the application ID. A screen that succeeded just after your client gave up then produces one result, not two.
The queue’s release rule must never depend on time or load. The Payoneer case is the warning: its backlog drained itself without review. A pending application leaves the queue in exactly one way, through a stored clear result or an analyst’s recorded decision. When the provider returns, drain the queue oldest first, at a controlled rate.
Tell the user that verification is taking longer than usual, when they will hear from you and how to reach support. Do not tell them they are approved pending checks, and there is no need to name the cause.
True hit: account hold, compliance lane, manual review against the official list
A hit puts the account on hold before a human sees it. A hold is your internal state, not yet a legal action: it freezes activation and outbound funds while an analyst decides whether the match is real. Hold first, because releasing a true match is a violation. Review second, because many alerts are false positives. OFAC’s FAQ 48 says screening software flags transactions that have nothing to do with OFAC targets, and that this is where human intervention becomes critical.
The analyst follows OFAC’s FAQ 5, updated on 9 September 2026. Compare every detail of the list entry (names and aliases, date of birth, nationality, identification numbers, addresses) with what the applicant provided, and gather more information if needed. If only a first name matches, or the date of birth differs, the match is probably not valid, and the analyst can release the hold and record why. If several details match, follow your procedure for likely matches: check whether a general or specific licence applies, then decide between blocking and rejecting. Persons on the SDN List and entities blocked under OFAC’s 50 Percent Rule are blocked. A sanctioned jurisdiction with no blocked person involved is a rejection.
Both outcomes carry paperwork. Reports to OFAC are due within 10 business days, and investigation records must be kept (31 CFR 501.601; ten years, according to OFAC’s introduction). OFAC states that it does not confirm potential matches or false positives, so the decision and its evidence are yours. Log who decided, when and why, together with the screening reference, the list entry and the data screened.
Two design choices from the Bank of America case belong in the queue logic. A rule that automatically suppressed alerts matching an earlier false-positive decision let payments to blocked banks through. A prioritisation flaw ranked exact name-and-date-of-birth matches as low or medium priority for an extended period. Never auto-dismiss on past decisions without re-checking the list entry, and send exact matches to the front of the queue.
How to test a sanctions screening API outage before it happens
Testing and auditing is one of the five essential components in OFAC’s 2019 Framework for OFAC Compliance Commitments. In the Bank of America release, OFAC says screening alert parameters and rules should be subject to audit and testing.
Start with a production-parity sandbox. A sandbox is a valid test environment only if it matches production in authentication, rate limits and response structure. If it accepts a key that production would refuse, or never rate-limits, the 401, 429 and timeout branches you test are not the ones you will meet.
Then break it on purpose. Run these as automated tests on every release, not as manual checks after deployment, and repeat them as a periodic drill:
- A revoked or expired API key: expect a 401, no activations and an alert within minutes.
- A malformed payload: the 400 is treated as a bug, not retried.
- A response slower than your client timeout: the account stays pending and the retry schedule starts.
- A 503 with a Retry-After header: the next attempt waits at least that long.
- A 200 OK with an empty body, truncated JSON or an unrecognised status value: not cleared.
- The provider returns after two hours with hundreds of queued applications: oldest first, and none released without a stored result.
- A watchlist feed that stops advancing: treated as an outage once it passes its freshness threshold.
- A hit: the account is held, a case is opened and nothing moves.
One regression test is worth more than the rest: it fails the build if any code path can set an account to active without a stored clear result.
Frequently asked questions
How long can a customer stay pending?
None of the OFAC guidance cited here sets a maximum pending period, so the limit is the one you write: the account stays inactive until a result exists. Set an SLA, escalate to compliance when it is breached, and decide the fallback in advance. One fallback is manual. OFAC publishes a free Sanctions List Search tool for single-name searches across its lists, so a trained analyst can clear priority applicants by hand during a long outage and record the result. It covers OFAC lists only, so it does not replace PEP or other list checks.
Can payments flow while screening is pending?
Not for a party with no screening result. Anything involving that party waits. OFAC’s FAQ 42 compares blocked property to a bouncing ball: once it comes into your possession, you must stop it. Customers who were cleared earlier are a different case, because their exposure is a list change after clearance. That is the monitoring question below.
What about ongoing monitoring outages?
Treat the downtime as a gap to close. When service resumes, re-screen the whole customer base against every list change since the last successful run, oldest first, and log the window. OFAC explains in its SDN List FAQs where to download every change to the list, which tells you exactly what a gap missed. Bank of America did this: after activating the February 2025 update on 12 February, it ran a lookback and self-disclosed the 26 transactions it had processed on 7 February.
Speed matters. The FATF glossary defines “without delay” for targeted financial sanctions as, ideally, within a matter of hours of a UN Security Council designation. National rules turn that into firm deadlines, so check yours, but a gap of days is the wrong order of magnitude.
