The Authoritarian API

How dictatorships learned to punish dissidents without touching them

On 30 October 2023, a Belarusian filmmaker named Andrei Hniot landed in Belgrade and was arrested at the airport.

He had not been arrested by Belarus. He had not been convicted of a crime in Serbia. He had been flagged by an Interpol Red Notice, a police cooperation request issued through an international system that treats criminal accusations as standardized inputs. Belarus had labeled him a fugitive on tax charges he denies. Serbia processed the alert. [1]

For a year his life was reorganized around a form. Seven months in Belgrade Central Prison, then house arrest with a tracking bracelet and one hour outside a day. The punishment was not only the possibility of extradition. It was the process itself, running through institutions that did not need to share Minsk’s politics to act on its signal.

In July 2024, Interpol deleted the notice, finding it incompatible with the provisions of its own constitution that bar the organization from political cases. [2]

The process did not stop.

Serbian proceedings ran three more months. An appeals court overturned the extradition order in September but sent the case back for retrial, and Hniot stayed under house arrest. He walked free on 31 October 2024, a year to the day after his arrest, because Serbian law caps pre-extradition detention at a year. The case was still open when he left the country. [1]

The accusation had been withdrawn at the source. The machinery kept running on momentum.

Authoritarian regimes have discovered that the most effective way to reach a dissident abroad is not always to send an agent. It is to submit a form. File the right paperwork, in the right format, through the right international channel, and the institutions of liberal democracy may do the rest. Border systems flag passports. Banks freeze accounts. Visa offices open reviews. Universities, employers and nonprofits hesitate.

The regime does not need to win. It only needs the process to run.

In the age of AI, that process is becoming faster, cheaper and harder to see. What authoritarian states will produce more of is not propaganda. It is machine-readable suspicion.

A blacklist, a terrorism label, a criminal charge, a state-media smear, a coordinated reporting campaign: in the pre-AI world these were political attacks, damaging reputations and creating bureaucratic friction. In the AI-mediated world they become data, absorbed into banking systems, visa screening, compliance software, hiring platforms, grant reviews and automated risk scores.

The regime no longer needs Western institutions to believe the accusation.

It only needs their systems to ingest it.

Procedural Trust

To understand the exploit, one must understand the system it exploits.

After World War II, liberal democracies built their international order around a sound moral intuition: power should be forced through procedure. Courts, treaties, banking rules, migration systems and police cooperation were designed to slow arbitrary force into channels that could be reviewed and contested. Procedure created a hidden vulnerability of its own.

Liberal institutions often assume that official-looking claims, even from adversaries, maintain some relationship to evidence and legal reasoning. A criminal charge is a criminal charge. A Red Notice is a Red Notice. The system is built to process such inputs and produce outputs: approvals, denials, delays, escalations.

Authoritarian regimes learned to exploit that assumption.

Freedom House has documented 1,375 direct physical incidents of governments reaching across borders to silence dissidents between 2014 and 2025: assassinations, abductions, illegal deportations, family intimidation. In at least eleven of the 2025 cases the perpetrator government used an Interpol notice to do it. [3]

But the most consequential shift may be less visible. It is administrative.

Authoritarian states learned to translate political hostility into bureaucratic form. A dissident does not need to be captured if he can be made institutionally toxic. A journalist does not need to be censored if she can be turned into a compliance problem.

The accusation becomes more powerful when it stops looking like propaganda and starts looking like a file.

That is the exploit: repression formatted for liberal processing.

Exported Blacklists

Mikhail Khodorkovsky, once Russia’s richest man and imprisoned by Putin for a decade before going into exile, described the mechanism with unusual precision in May 2026, testifying before the European Parliament’s special committee on the European Democracy Shield. [4]

He and his colleagues in the Russian Anti-War Committee have been designated terrorists by the Kremlin. Most European states, he told the committee, understand this as a political decision with nothing to do with security. In practice it means that on every trip, and in every interaction with a bank or migration authority, there is a terrorism flag in his file. For thousands of less prominent exiles the consequences are duller and heavier: banks close accounts because compliance departments fear any link to a name marked suspicious, migration authorities delay status extensions because they cannot interpret a Russian accusation, passports go unrenewed until people are stateless in practice. Europe’s financial system, he argued, has become an unwitting instrument of the repression it opposes.

A Russian blacklist becomes a European banking problem. A Kremlin terrorism label becomes a migration delay. A politically motivated designation becomes a compliance risk, not because the institution endorses the claim, but because the claim has entered its data environment and triggered its risk logic.

This is not traditional censorship. It is repression through interoperability.

Igor Pestrikov’s case shows the shape it takes. A Russian businessman and shareholder in a magnesium and rare-earths producer, he left in 2022 after refusing to supply metal to buyers the state had designated, and settled in southern France. Russia did not publish a Red Notice against him. It used a diffusion, a request circulated directly between national police bureaus, which moves faster and draws less scrutiny than a notice the target can contest. For the two years it was live his bank accounts were frozen and he could not rent an apartment. Interpol’s review commission eventually deleted it as politically motivated. [5]

No bank officer accused him of anything. A screening system returned adverse results and institutions declined the exposure.

The pattern repeats wherever a risk system sits between a person and a service. No one has to say: we believe the dictatorship. Each says only: this is complicated.

In a risk-managed institution, complicated often means not worth it.

The authoritarian state creates the label. The liberal system supplies the enforcement surface. The punishment arrives through the administrative caution of institutions that would never endorse the politics behind it.

The API

The API metaphor is not decorative. An application programming interface is a standard way for one system to send instructions to another: an input format, a receiving system, an execution layer. Send the right request, in the right format, and the receiving system acts.

Authoritarian states have learned to exploit all three. The input format is the official-looking signal: a criminal charge, a terrorism designation, an Interpol request, a state-media article. The receiving system is liberal institutional infrastructure: banks, visa offices, compliance vendors, platforms, universities, grant makers. The execution layer is not a police order. It is a risk decision: delay the visa, close the account, freeze the grant, restrict the profile.

The volume is rising. Interpol issued 15,548 Red Notices in 2024, a 27 percent increase and the largest annual total on record, while its own compliance task force refused or cancelled 2,462 notices and diffusions that year, up 54 percent, also a record. [6] More inputs arrive and more bad ones are caught, but the catching happens after the processing has begun.

Files given to BBC World Service and Disclose by a whistleblower, published in January 2026, sharpened the picture. Interpol’s review commission removed at least 322 notices in 2024 as unjustified, and Russia drew more complaints to it than any other member state. Fewer than one in ten of roughly 86,000 active Red Notices are published at all, and measures constraining Russian abuse were quietly relaxed during 2025. [7]

Hniot’s notice was corrected. The harder question is how much punishment lands before a correction arrives, and how little it undoes.

A secure API assumes some inputs are hostile and validates accordingly. Liberal democracy still runs too many systems as if official-looking inputs are valid by default. That assumption once made procedure possible. Now it makes procedure exploitable.

Risk Laundering

Banks, payment processors and compliance vendors are not villains in this story. They are operating rationally inside a system never designed to distinguish political persecution from genuine risk at scale. Risk systems are not the problem. They are being asked to process signals from regimes that manufacture risk as a political weapon.

Authoritarian regimes do not need to invent the risk system. They only need to contaminate it. The Financial Action Task Force, which sets global standards for fighting financial crime, opened a review of its own unintended consequences in February 2021. The stocktake it published that October sorted the damage under four headings: de-risking, financial exclusion, undue targeting of nonprofits, and curtailment of human rights, the last focused on due process and procedural rights. [8]

A standards body examined its own rules and found they were eroding procedure. The diagnosis came from the institution that wrote the standard.

De-risking is the named mechanism: banks withdrawing services from whole categories of customer rather than pricing the risk of any individual one. The categories that absorbed the damage were nonprofits, diaspora populations and people in conflict-affected regions, which is to say the populations transnational repression selects for. [8]

A bank does not have to conclude that an activist is guilty. It only has to conclude that she is operationally complicated.

This is what risk management can launder. A Kremlin designation need not be cited directly if it has already shaped the data environment around the person. A Chinese state-media smear need not be believed if it reappears downstream as adverse media. The machine does not repeat the lie. It operates the uncertainty the lie creates.

Machine Suspicion

AI does not create this exploit. It changes its physics.

The pre-AI version depended on bureaucratic friction. A blacklist had to be noticed by a compliance officer, translated, escalated and interpreted by someone with enough context to evaluate it. Each step was a chance for a person to pause or override. AI removes many of those pauses.

The signals Khodorkovsky described also stayed fragmented: a criminal file in one system, state-media coverage in another, an unrenewed passport in a third. In the AI-mediated world the fragments become legible to one another. Name-matching connects aliases across languages. Compliance software flags unresolved allegations.

Adverse-media screening is where this concentrates. LSEG’s World-Check media tool applies machine learning and automated tagging across more than thirteen thousand vetted sources in twenty-four languages, clusters the results into discrete events, and delivers them to compliance platforms through an API. [9] At that point the metaphor stops being a metaphor.

Nor is the automation merely a commercial preference. Successive European anti-money-laundering directives widened the range of firms required to screen open-source media and pushed those checks toward automation. The pipeline is built, mandated and running, and whatever an authoritarian jurisdiction produces in text enters it as text. [10]

No institution has to believe the regime. Each has only to downgrade the person by one notch: enhanced review, delayed approval, account closure, grant hesitation, reputational caution.

A model does not say: this dissident is a terrorist. It says: elevated risk, unresolved allegations, adverse media present, enhanced review recommended. Not belief. Probabilistic suspicion. And probabilistic suspicion, distributed across enough institutions, produces political outcomes without any single institution making a political decision.

An authoritarian lie no longer has to survive as a lie. It can survive as a probability.

Hostile Inputs

There is an obvious objection: not every criminal case from an authoritarian state is fabricated. Not every adverse-media flag is propaganda. Democracies cannot ignore all foreign-origin risk signals.

The answer is not blanket dismissal. It is adversarial weighting.

A terrorism designation from an independent judiciary and a court filing from a captured legal system should not enter downstream risk systems at the same default weight. Cybersecurity already runs on this principle: a login attempt from a known malicious network is handled differently from one on a trusted device. A secure system does not abolish communication. It validates the source, checks the payload and limits the damage any single input can cause.

The line is not always clean. In May 2025, El Salvador obtained Red Notices against two Salvadoran lawyers living in Spain, Ivania Cruz and Rudy Joya. UN experts identified them as retaliation tied directly to the lawyers’ human rights work, and Interpol revoked them as politically motivated. The two still faced extradition proceedings in Spain and did not secure asylum until 2026. [11]

A scheme keyed to a short list of usual suspects would have missed that case, and revocation, once again, did not stop the process the notice had started. The design problem is harder than sorting regimes into columns. What has to be weighted is the quality of the producing system: whether its judiciary can be overruled by its executive, whether a designation carries reasons, whether the target can see the accusation and answer it.

The goal is not to make dissidents untouchable, but to stop authoritarian regimes turning liberal procedure into an enforcement arm.

Patching Procedure

Fixing this does not require dismantling procedure. It requires making procedure adversary-aware. Five reforms would help.

Provenance labeling. Risk signals from regimes with documented records of political persecution should carry origin flags before entering compliance, migration, employment and grant systems. The source of an allegation should travel with the allegation.

Corroboration requirements. High-consequence decisions such as account closures, visa denials, grant cancellations and platform removals should not execute on a single authoritarian-origin signal without independent corroboration.

Source auditability. Compliance vendors should disclose whether their adverse-media summaries draw on state media, regime-linked outlets or opaque source chains in illiberal jurisdictions. Provenance is the foundation of judgment, not a technical footnote beneath it.

Fast-track review. Banks, platforms, payment systems and immigration offices need escalation paths for people plausibly targeted by transnational repression, separate from the standard commercial dispute queue.

Meaningful appeal rights. Anyone denied banking, visas, grants or essential services by an automated risk system should receive human review and a substantive explanation.

That last reform is the one most likely to fail, and the failure mode deserves naming. Human review placed above an automated judgment frequently works as a liability buffer rather than a check, supplying a face for the decision without supplying judgment about it. A reviewer who cannot see the source of the flag, cannot overrule the model and has no time to read the file is not a safeguard. He is a signature. Appeal rights are worth something only where the reviewer can see the provenance the first reform requires, can rule against the system, and leaves a record for doing so. Otherwise the appeal is one more piece of procedure the exploit can run through.

These reforms do not ask institutions to abandon caution. They ask them to become cautious about the right thing. The risk is not only that a dangerous person slips through. It is that a dictatorship learns how to make innocent people administratively radioactive.

The Process Runs

The postwar liberal order was built around trust in process. Authoritarian regimes learned to weaponize that trust, and AI is turning the exploit into infrastructure.

The exiled journalist does not disappear because one institution makes a tyrannical decision. She disappears administratively because many institutions make cautious ones. A dissident does not need to be convicted if he can be risk-scored. A journalist does not need to be censored if she can be deplatformed by procedure. An exile does not need to be imprisoned if his identity, banking, travel, work and reputation can be quietly degraded across systems that all claim neutrality.

The old danger was propaganda: a lie people might believe.

The new danger is administrative: a lie systems might process.

The future of authoritarian power may not always look like a knock on the door. It may look like an automated decision no human owns, based on a label no one verifies, produced by a regime no one admits they obey.

Democratic institutions were built to fight force. They are less prepared to fight formatted suspicion. The authoritarian API works not because the lie is believed, but because the lie gets processed. Procedure, without adversarial awareness, will keep running it.

Notes

[1] PEN International, case file “Andrej Hniot,” and “Serbia: Belarusian filmmaker and journalist Andrej Hniot to be retried,” December 2024: the arrest on 30 October 2023, seven months in Belgrade Central Prison, transfer to house arrest on 5 June 2024, and the Court of Appeal’s order of 11 September 2024 sending the case back for retrial. On the release and the statutory one-year ceiling on pre-extradition detention, see Associated Press reporting of 1 November 2024 and Balkan Insight, “Belarus Activist Freed From Detention in Serbia Leaves for EU,” 1 November 2024, quoting defense counsel Filip Sofijanic that proceedings remained pending and that Belgrade’s High Court was still awaiting documentation from Belarus when Hniot departed.

[2] Pozirk, “Interpol deletes red notice for filmmaker Hniot,” 19 July 2024, reporting the Belarusian Association of Journalists’ account of the Interpol General Secretariat’s letter to the EU delegation in Serbia, which found the notice inconsistent with Articles 2 and 3 of the Interpol Constitution.

[3] Freedom House, Collaboration and Resistance: Tracking Transnational Repression in 2025, April 2026. The database records 1,375 direct, physical incidents by 54 governments across 107 host countries between 2014 and 2025, including 126 new incidents in 2025. Detention (49) and unlawful deportation (48) were the most common tactics, and in at least 11 such cases perpetrator governments used Interpol notices, which the report reads as evidence that Interpol’s reforms have not yet closed the avenues available to member governments.

[4] Mikhail Khodorkovsky, testimony to the European Parliament’s Special Committee on the European Democracy Shield, 5 May 2026, published as “Khodorkovsky in the European Parliament: How the Kremlin Exports Repression to Europe.” Khodorkovsky also asked the committee to recommend guidance on banking de-risking, developed with DG FISMA and national regulators, obliging banks to distinguish genuine security threats from politically motivated accusations by authoritarian states.

[5] BBC World Service and Disclose, joint investigation published 26 January 2026. Pestrikov’s account, given publicly for the first time, describes frozen accounts and refused tenancy across the two years the Russian request was active, and its eventual deletion by the Commission for the Control of INTERPOL’s Files. On the distinction between notices and diffusions, the same reporting describes member states using Interpol’s messaging channels to trace people abroad in place of a notice the subject can challenge.

[6] INTERPOL, Annual Report 2024; see also Red Notice Monitor, “Increasing number of Red Notices issued,” 2025. 15,548 Red Notices published in 2024, a 27 percent rise and the highest annual total on record; 2,462 notices and diffusions refused or cancelled by the Notices and Diffusions Task Force, against 1,598 in 2023.

[7] BBC World Service and Disclose, 26 January 2026; and Amnesty International, “Global: Misuse of Interpol red notices to target dissidents a grave institutional failure,” 26 January 2026. The figure of at least 322 removals by the Commission for the Control of INTERPOL’s Files in 2024, and Russia’s position at the head of complaints to that commission, derive from the leaked files. Disclose reports that fewer than 10 percent of approximately 86,000 active Red Notices have been disclosed publicly, and that as of September 2024 the largest holders were Russia (4,817), Peru (4,457) and Tajikistan (3,493).

[8] Financial Action Task Force, High-Level Synopsis of the Stocktake of the Unintended Consequences of the FATF Standards, 27 October 2021. The project was launched in February 2021 and examines four themes: de-risking, financial exclusion, undue targeting of NPOs, and curtailment of human rights with a focus on due process and procedural rights. Recommendation 8, governing the treatment of non-profit organizations, was revised in 2023.

[9] LSEG, product documentation for World-Check One Media Check, describing machine learning and intelligent tagging applied to more than 13,000 vetted sources across 24 languages, clustering of content into discrete events, and delivery by platform or API.

[10] The Fifth Anti-Money Laundering Directive widened the range of obliged entities required to conduct open-source media checks and introduced automation requirements for adverse-media screening; the Sixth added predicate offences that extended the screening scope further. See also European Center for Not-for-Profit Law, How AI Is Powering Transnational Repression, 2026, a scoping study covering algorithmic risk assessment, automated blacklisting and AI-enabled surveillance in the transnational-repression context.

[11] Freedom House, Collaboration and Resistance: Tracking Transnational Repression in 2025, April 2026, recording the May 2025 Red Notices against Ivania Cruz and Rudy Joya, the UN experts’ finding of a direct connection to their human rights work, Interpol’s revocation on political-motivation grounds, and the continuation of Spanish extradition proceedings until asylum was granted in 2026.