When the fabrication becomes the weapon.
Synthetic media generation now lets belligerents corrupt the records, communications, and identifications that make a captured person legible as a human being owed dignity. MAESTRO4IHL maps where generative AI threatens prisoner-of-war protections — so the harm becomes foreseeable, and therefore preventable.
From Hague law to Geneva law
The AI-in-conflict debate has matured around the conduct of war. It has been near-silent on how AI renders war more inhumane.
Reorient the discourse
Move from Hague-law concerns about the conduct of hostilities toward Geneva-law concerns: the protection of individual human dignity in captivity.
Adapt a threat model
MAESTRO4IHL adapts the MAESTRO agentic-AI framework to assess the harm surface that synthetic media create for POW entitlements under GCIII.
Lay the groundwork
Found future technical and policy measures — this platform, red-teaming exercises, and governance mechanisms — to address the vulnerability before harms occur.
Just as much as the laws of war must become a political priority, so too must they become a sociotechnical priority on the part of all AI actors. — On the case for a sociotechnical turn in red-teaming AI systems in armed conflict
A fragile information regime
Why POW protection depends on records, contact, and identity staying valid, accessible, and attributable.
The MAESTRO4IHL framework
Seven layers, six IHL adaptations, and six analytic steps for mapping where synthetic media threatens entitlements.
Walk a scenario
Step through three worked cases as a responsible actor: decompose, map threats, trace the chain, mitigate.
Who defends each layer
A distributed-defense matrix that places the burden where capacity sits, and never on the victims.
The work behind it
The peer-reviewed basis, the sociotechnical-turn argument, and where the project goes next.
Bring it to your work
For AI labs, humanitarian institutions, and policymakers ready to pressure-test these protections.
The POW information regime
POW protections do not depend on a single record. They depend on a fragile, distributed information regime that must satisfy three conditions to function at all.
Validity
Information about POWs — records, communications, identifications — corresponds to reality.
Access
Information can be readily reached by the right stakeholders: families, home and detaining states, the ICRC, and the prisoner themselves.
Attributability
Information carries clear provenance, allowing it to be traced back to a legitimate actor.
The MAESTRO4IHL framework
MAESTRO's seven-layer architecture maps cleanly onto the pipeline that produces POW-targeted synthetic media — data → model → orchestration → deployment → ecosystem. Six targeted adaptations re-anchor it on humanitarian outcomes and IHL.
Scenario explorer
Select a scenario, then step through the framework as a responsible actor would — decomposing the system, mapping threats across layers, tracing the attack chain, and assigning mitigations.
Who is the responsible actor?
Decomposing the regime into layers reveals who is best positioned to defend each one. Responsibility is shared and continuous — and it never falls on the victims of harm.
The work behind the project
The AIHL Project rests on a straightforward argument: the protection of people hors de combat deserves the same sociotechnical scrutiny the field already gives to the conduct of hostilities.
A sociotechnical turn for the laws of war
Debate over artificial intelligence in armed conflict has matured around Hague law — targeting, proportionality, weapons review, escalation. It has been near-silent on how the same technologies make war more inhumane for those already in custody.
This work reorients that debate toward Geneva law: the dignity, contact, and accurate record-keeping owed to prisoners of war. It shows how generative synthetic media can corrupt the fragile information regime on which POW protection depends, and adapts the MAESTRO agentic-AI threat-modeling framework into MAESTRO4IHL, an instrument that maps the harm surface across seven layers and six analytic steps.
Three worked scenarios — a propaganda deepfake, falsified capture cards, and an audio-cloned call to a prisoner's family — demonstrate the model, and a distributed-responsibility analysis assigns defensive obligations in proportion to capacity, never to the victims. MAESTRO4IHL's central demand is that AI actors treat the laws of war as a sociotechnical priority, not only a political one.
6 analytic steps · 3 worked scenarios
Grounded in GCIII & customary IHL
Scenario library
An openly navigable, growing library of worked cases, expanding as new synthetic-media threats are analysed.
Red-teaming exercises
Operationalizing MAESTRO4IHL as adversarial evaluation against frontier models, with structured probes and scoring.
Governance mechanisms
Folding synthetic-media harms into IHL governance, accountability, and cross-actor reporting before harms occur.
Who builds this, and why
The AIHL Project is the work of two researchers who saw the same gap from different vantage points and decided it was worth naming.
We formed the AIHL Project to bring attention to a growing but under-resourced threat area: synthetic media attacks with implications across international humanitarian law. The generative tools that can corrupt the records, communications, and identifications protecting people in conflict reach into many IHL areas, wherever a protected person has to stay legible and their record has to stay trustworthy. MAESTRO4IHL is the Project’s threat model for the surface we have worked first and most closely: synthetic-media harms to prisoner-of-war protections under GCIII. The capability to cause this harm already exists. Building the defensive attention to match it is the work this project exists to do.
Nathan Heath
Nathan Heath is a decision scientist and AI safety researcher with 14 years of experience working at the intersection of geopolitics and emerging technology. He serves as the Founder & CEO of Syntony, where he leads efforts in adversarial evaluation, governance architecture, strategic risk advisory, and software development for organizations operating high-stakes systems.
He previously served as a Senior Research Scientist at National Security Innovations, where he led work on emerging technology risk, AI integration, and security cooperation for U.S. government clients, examining the interconnected challenges posed by emerging technologies and the implications of evolving geopolitical alliances for strategic and operational environments.
Nathan also serves as a Red Teamer for OpenAI and Anthropic, where he conducts adversarial evaluation of frontier AI systems. He is additionally an Expert Advisor to the Cloud Security Alliance, a Security Fellow with the Truman National Security Project, a Member of MIT AI Alignment, and a Contributing Researcher with the Oxford Martin AI Governance Initiative.
Nathan’s current academic research focuses on the application of systems dynamics modeling to frontier AI risk, methods for improving defenses of testimony archives against synthetic media attacks, and defense industrial cooperation in Europe.
Nathan holds an M.A. in Law and Diplomacy from The Fletcher School at Tufts University and also studied Politics, Philosophy, & Economics (PPE) at Oxford and diplomatic and business strategy at Harvard. His work has been presented at the International Association for Safe and Ethical AI (UNESCO House), UNIDIR, the Cambridge Centre for Geopolitics, the Minderoo Centre for Technology & Democracy, the American University of Paris, the International Institute for Justice, and the UK MoD Deterrence and Assurance Academic Alliance, and has appeared in War on the Rocks, RAND Europe, World Politics Review, PRISM, and The Washington Post.
Nathan writes A Mind of Their Own, a Substack exploring AI consciousness, governance, and safety, and regularly contributes commentary on AI safety and analytic philosophy to LessWrong. He also leads the North Carolina chapter of the OpenAI Forum and contributes research to the Meta Oversight Board and the EvalEval Coalition.
Wm. Matthew Kennedy, PhD
Wm. Matthew (Matt) Kennedy is a Marie Skłodowska-Curie Postdoctoral Fellow.
He researches AI ethics, sociotechnical safety, and political economy, focusing on three themes in particular: AI evaluation and red-teaming methodologies; AI and knowledge production, with special interest in AI in education, epistemic harms, and interdisciplinarity in AI; and decolonial AI.
He is also an affiliate of several institutions and organizations, including the University of Sydney Centre for AI, Trust, and Governance, the Machine Learning & Society team at Women at the Table (a systems change nonprofit based in Geneva), the AI and Education at Oxford University research hub (AIEOU), and the Center for AI and Digital Policy, as a research group member. He contributes regularly to events put on by Humane Intelligence (a nonprofit with the mission of building open AI assurance tools and community) and is an inaugural ORF US-India AI Fellow.
Before joining the OII, he received his PhD from the University of Sydney in the history of colonialism, international law, and “scientific” governance. Thereafter, he held research fellowships at the University of Texas at Austin and the University of Sussex. After a period of time spent in the technology industry, he became increasingly involved in AI evaluations functions, eventually leading a small team tasked with structuring evaluations of AI systems, collaborating with external colleagues to produce original research on AI ethics, safety, and governance, and occasionally advising U.S. and international public officials on matters of AI policy.
While at the OII, he will begin a new project that investigates the colonial past as a source of sociotechnical foresight for guiding the design, development, deployment, use, and governance of AI systems.
Pressure-test the protections, before they fail
MAESTRO4IHL is built to be used. It is most valuable in the hands of the actors best positioned to find and close these gaps in their own systems and mandates.
The harm surface here is foreseeable, which means it is preventable. Whether you build the models, safeguard the platforms, hold the humanitarian mandate, or write the rules, MAESTRO4IHL gives you a shared language for naming where synthetic media threatens prisoner-of-war protections — and a method for deciding what to do about it.
Harden generative pipelines against the jailbreaks and misuse pathways the scenarios trace.
Detect and interdict synthetic POW media before it reaches families, states, and the record.
Anchor evaluation on humanitarian outcomes and escalate failures of the information regime.
Bring record-integrity and dignity harms explicitly into IHL governance and accountability.
Scope
A briefing maps MAESTRO4IHL onto your systems, mandate, or portfolio and identifies the layers most exposed.
Evaluate
Worked scenarios and red-teaming exercises pressure-test those layers against realistic adversaries.
Operationalize
Findings become concrete mitigations and monitoring, assigned in proportion to your capacity.
Bring MAESTRO4IHL to your organization.
Request a briefing, propose a red-teaming engagement, or talk through how the framework applies to your work.