There is a line in the architecture of modern warfare that military ethicists, legal scholars, and arms control advocates have spent years trying to define, codify, and protect. It is the line between a weapon that kills on human command and a weapon that kills on its own judgment. The former has been the subject of centuries of international humanitarian law. The latter has no legal framework, no treaty prohibition, and, as of March 2026, no technological barrier. NORDA, a defense AI company that Forbes has now confirmed holds active contracts with Ukrainian military forces and U.S. contractors operating in the Iran theater, is the company that may have crossed that line first.
The Forbes investigation, based on documents obtained from three sources with direct knowledge of NORDA's programs, government procurement records, and interviews with former employees, reveals that NORDA has developed an autonomous targeting and engagement system that can identify, classify, and engage targets without human authorization once a mission is activated. The system has been tested in field conditions, has been deployed in a limited capacity in eastern Ukraine, and is under evaluation for use in the Iran theater, where U.S. and allied forces have maintained a persistent drone presence since the escalation of 2024.
What NORDA Is and Where It Came From
NORDA is a defense AI company that most people outside the defense technology community have not heard of. It is not Palantir, Anduril, or Shield AI. It does not have a public-facing communications strategy, does not attend commercial technology conferences, and has never issued a press release. Its name does not appear in standard defense contractor databases because it operates primarily through subcontracting relationships with prime contractors who hold the direct government contracts.
The company was founded in 2019 by a team of computer vision engineers who had previously worked at a classified AI program within the National Geospatial-Intelligence Agency. Its founding thesis was that the primary bottleneck in unmanned aerial system performance was not the hardware, not the communications, and not the human operator skills, but the latency introduced by the human decision loop. A drone that must transmit video to a human operator, wait for the operator to assess a target, and then receive an engagement command can only engage targets that remain stationary or slow-moving during that decision cycle. Against fast-moving, time-sensitive targets, the human-in-the-loop requirement is, from a purely operational standpoint, a significant tactical limitation.
NORDA's founders believed they could eliminate that latency by replacing the human decision step with an AI classification and engagement system that operated at machine speed. The technical challenge was not primarily one of capability: AI systems capable of identifying human figures, military vehicles, and weapons had existed in laboratory conditions since at least 2015. The challenge was deploying such systems at the reliability levels required for lethal autonomous operation in complex real-world environments, where the consequences of misidentification are measured in human lives.
How the System Works: From Target Selection to Terminal Guidance
Understanding what NORDA has built requires understanding the spectrum of autonomy that currently exists in unmanned aerial systems, because NORDA's system is not a binary shift from human-controlled to fully autonomous but a specific configuration of that spectrum that is new in its operational implementation.
Current state-of-the-art semi-autonomous drone systems, including the most advanced Ukrainian FPV systems and the U.S. military's current generation of loitering munitions, operate on a model where the human selects the target, the AI guides the terminal engagement. The human operator views a target area through the drone's camera, designates a specific point or object as the aim point, and releases control to the AI guidance system, which executes the final intercept trajectory. The human has made the kill decision. The AI has executed it. This is the standard for current state-of-practice and is explicitly compliant with DoD Directive 3000.09, which requires meaningful human control over lethal force decisions.
NORDA's system changes the model at the most consequential step. In NORDA's architecture, the human operator sets mission parameters before the drone is launched: authorized target categories, engagement zone boundaries, rules of engagement criteria, and mission duration. Once the drone is airborne and has entered the designated engagement zone, the AI system takes over completely. It scans the environment using a multi-modal sensor suite combining optical cameras, thermal imaging, and millimeter-wave radar. It applies a classification algorithm trained on an extensive dataset of military-relevant objects to identify potential targets within its engagement parameters. When a target meeting the defined criteria is identified with sufficient confidence, the system autonomously executes the engagement without further human input.
The confidence threshold is configurable. At higher confidence thresholds, the system will engage fewer targets but with greater accuracy. At lower thresholds, it will engage more frequently with a higher risk of misidentification. The choice of threshold is a human decision made before deployment, but the targeting and engagement decisions during operation are entirely machine-driven.
The AI Classification Pipeline
NORDA's target classification pipeline is built on a transformer-based neural network architecture trained on a dataset that former employees describe as "multi-terabyte scale," comprising labeled imagery from multiple conflict zones including Ukraine, Syria, and Yemen. The network is fine-tuned specifically for the object categories relevant to NORDA's operational contracts: military vehicles, weapons systems, and, critically, armed combatants.
The classification of armed combatants is the most technically and ethically fraught element of the system. The network is trained to distinguish between civilians and combatants based on visual cues including weapons visibility, body posture, movement patterns, and association with known military equipment. This is a classification task that human soldiers with extensive training struggle to perform reliably in real combat environments. Whether an AI system can perform it reliably enough to authorize lethal force autonomously is the central question that NORDA's operational deployment will answer in the field rather than in a laboratory.
Battle-Tested Partial Autonomy: What Has Already Happened
NORDA's current partial autonomy deployment in Ukraine is not an experiment. It is an operational system that has been used in the eastern sectors of the front since mid-2025. In this deployment, NORDA's system operates in the mode described above: human target selection, AI terminal guidance. The operational results have been significant enough to generate substantial interest from Ukrainian military procurement offices and from U.S. defense officials monitoring the conflict.
The primary operational advantage demonstrated in the Ukrainian deployment is engagement speed. In a conventional teleoperated FPV drone engagement against a moving military vehicle, the average time from target identification to impact, accounting for human recognition, designation, and guidance, is approximately 8 to 12 seconds. NORDA's AI-guided terminal system reduces the time from designation to impact to under 2 seconds, and the guidance precision against moving targets is substantially higher than human-guided systems at comparable drone speeds.
The secondary advantage is electronic warfare resilience. Teleoperated drones are vulnerable to communications jamming: if the radio link between the operator and the drone is disrupted, the drone loses guidance. NORDA's autonomous terminal guidance operates independently of the communications link; once the engagement is initiated, jamming cannot prevent it. This is a significant tactical advantage in the Ukrainian environment, where Russian electronic warfare systems are specifically designed to disrupt Ukrainian drone operations.
The Next Step: Full Autonomy Without Human-in-the-Loop
The Forbes investigation's most significant revelation is not NORDA's current deployment but what the company is developing for its next generation system. Internal documents obtained by Forbes describe a program internally designated as AUTONOMOUS SENTINEL, which removes the human target designation step entirely. In AUTONOMOUS SENTINEL, the drone is given a mission package before launch, flies to the designated area, applies its classification pipeline autonomously, and engages qualifying targets without any human interaction during the mission.
The technical barriers to AUTONOMOUS SENTINEL are primarily confidence and reliability. NORDA's engineers have, according to the Forbes documents, achieved classification accuracy rates for military vehicle targets exceeding 98.5 percent in controlled test environments. For armed combatant targets, the accuracy in test environments is lower, at approximately 94 percent, and the performance in complex real-world environments with poor lighting, obscuration, and adversarial countermeasures is not publicly documented.
A 94 percent accuracy rate sounds impressive in an abstract statistical sense. In the context of lethal force, it means that for every 100 engagement decisions the system makes, 6 may be misidentifications. At scale, across a fleet of hundreds of autonomous drones operating in a contested urban or semi-urban environment, that error rate translates to a significant number of potentially unlawful killings. This is the mathematical reality that underpins the ethical debate about autonomous weapons, and NORDA's development program is confronting it in operational conditions rather than in the international forums where that debate is supposed to be resolved.
No existing international treaty explicitly prohibits fully autonomous lethal weapons systems. The Convention on Certain Conventional Weapons' Group of Governmental Experts has debated the issue since 2014 without reaching binding conclusions. Under current U.S. law, there is no statutory prohibition on deploying autonomous lethal systems. DoD Directive 3000.09 requires "appropriate levels of human judgment" but does not define what that means in the context of pre-programmed engagement criteria.
The Iran Theater: First Autonomous Drone Warfare in a U.S. Conflict?
The Ukraine deployment, while significant, operates in a context where U.S. involvement is indirect: American-supplied technology used by Ukrainian forces in a conflict where the United States is not a direct belligerent. The Iran theater is different. Since the events of late 2024 that led to a sustained U.S. and allied drone campaign targeting Iranian military infrastructure and Revolutionary Guard facilities, U.S. contractors and intelligence agencies have maintained a persistent unmanned aerial presence over and around Iranian territory.
Forbes sources indicate that NORDA has been contracted by at least one U.S. government agency, identity not confirmed, to evaluate AUTONOMOUS SENTINEL's feasibility for use in the Iran theater. The evaluation is described as ongoing and pre-decisional, meaning no commitment to deployment has been made. But the existence of the evaluation, in a theater where U.S. forces are an active party, raises a question that the semi-autonomous Ukraine deployment does not: is the United States prepared to authorize the first fully autonomous lethal drone engagement in a conflict where Americans bear direct legal and moral accountability?
The legal implications are significant. Under the law of armed conflict, parties to a conflict bear responsibility for the weapons they use and the targeting decisions those weapons make. If a fully autonomous system makes an erroneous targeting decision that results in civilian casualties, the question of legal accountability becomes enormously complex. Is the manufacturer responsible? The military commander who authorized deployment? The operator who set the mission parameters? Current international humanitarian law was not designed to answer these questions, and the United States has not publicly articulated how it would assign accountability for autonomous weapon engagement errors.
Comparison to Israel's Autonomous Systems
The most instructive comparison to NORDA's development program is Israel's, because Israel is the country that has advanced furthest in the operational deployment of AI-assisted and autonomous targeting systems. The Lavender and Gospel AI systems, which generated international controversy following their reported use in Gaza, represent one point on the spectrum: AI systems that generate target recommendations for human approval. The Harop loitering munition, which autonomously hunts radar emissions, represents another: full autonomy in a narrow sensor-to-shooter cycle with no human decision required once the weapon is launched.
Israel has not publicly confirmed the development of fully autonomous human-targeting systems, but defense researchers who have studied the country's autonomous weapons procurement patterns note a clear trajectory toward increased automation at every stage of the kill chain. Israeli defense companies including Elbit Systems, Rafael, and Israel Aerospace Industries have all announced AI targeting systems with increasing levels of autonomous classification capability. The gap between "AI-recommended target with human approval" and "AI-selected target without human approval" is being closed incrementally, in public and in private, by multiple Israeli programs simultaneously.
NORDA's significance is that it represents an American company following a similar trajectory, but in a commercial rather than state-sponsored context. Israel's autonomous systems are government programs, developed within a framework of state accountability, however imperfect. NORDA is a private company developing autonomous lethal systems under commercial contracts, with a governance framework that is entirely voluntary and a regulatory environment that provides no binding constraints. This is a fundamentally different accountability structure, and it is the structure that may produce the first fully autonomous lethal drone engagement in a U.S.-connected conflict.
"The question is not whether autonomous weapons will be used. They already are, in partial form. The question is who gets to decide where the line is drawn, and whether that line will be drawn at all."
-- Former DARPA Program Manager, speaking on background, March 2026
What Happens When Both Sides Have Autonomous Drones
The strategic implications of mutual autonomous drone deployment have received surprisingly little analytical attention given their potential significance. The analysis that does exist, primarily from the think tanks and war colleges that model future conflict scenarios, suggests that bilateral autonomous drone deployment introduces instability dynamics that have no historical precedent.
In a conventional conflict with human-in-the-loop weapons, the tempo of engagement is constrained by human decision-making speed. Escalation requires human decisions at each step, and those decisions take time: time to assess the situation, consult with commanders, evaluate alternatives. That temporal friction is a structural feature of human-controlled warfare that has historically provided opportunities for de-escalation, cease-fire negotiations, and conflict termination.
In a conflict where both sides have autonomous drones operating at machine speed, that temporal friction disappears. An autonomous drone strike by one side can be detected, classified, and responded to by the other side's autonomous systems in seconds. Escalation cycles that would take hours or days in conventional warfare could complete in minutes. The risk of unintended escalation to catastrophic levels, including the use of weapons of mass destruction if either side interprets autonomous drone activity as a precursor to a larger attack, is substantially higher in an autonomous-versus-autonomous conflict.
This is not a hypothetical concern. Russia has deployed autonomous strike drones in Ukraine. Ukraine is deploying increasingly autonomous systems in response. Both sides are operating systems that make targeting decisions faster than human oversight can track. The current conflict has not yet experienced a major autonomous system engagement error that sparked escalation, but the statistical probability of such an error increases with every additional autonomous system deployed and every additional hour of autonomous operations conducted. The question is not whether it will happen but whether the international community will have established any framework for managing it when it does.
NORDA's development program is not the cause of this risk. It is a symptom of a technological trajectory that no single actor controls and that no existing governance framework has the tools to constrain. The day drones stop needing humans may already have arrived. The day humanity decides what to do about that has not.