EXCLUSIVE Artificial intelligence, the new weapon of Washington. How Claude transformed from a means of attack into a source of discord between the Pentagon, Trump, and Anthropic.

The integration of the Claude model, developed by Anthropic, in the selection of targets for the airstrike on Iran marks the shift of commercial artificial intelligence from the peripheral support area to the center of the operational architecture of war. The use of a generalist model, initially intended for civilian use, to assist in identifying and prioritizing targets in a coordinated Pentagon offensive indicates that AI is becoming a structural component of force projection, not just an auxiliary tool.

How Claude works in target selection

In the classic scheme, target selection is based on the fusion of intelligence information, images, intercepts, and contextual assessments, all filtered through successive layers of human analysis. With the introduction of Claude, this flow is reconfigured: the model becomes the main aggregator of data, the system that condenses a massive volume of information into a coherent list of targets and scenarios, each with an implicit profile of relevance and risk. ​ The military decision-maker no longer works with "raw" reality, but with an algorithmic representation of it. Additionally, the use of Claude during the execution of the attack, for updating targets based on the evolution of the situation, means that the model is present in the tactical loop, not just in the planning phase. Any inference error by the system can have immediate consequences on target selection and, implicitly, on casualties.

"Invisible pre-selection" and the change in the decision-making framework

An essential effect of this type of integration is the emergence of an "invisible pre-selection." Numerous hypotheses, contexts, and targets can be excluded from human debate for the simple reason that the model does not consider them significant enough to bring to the forefront. What effectively reaches the decision-maker's table is the product of an algorithmic filtering process, the criteria of which are not transparent in real-time. ​

At the same time, the recurrent use of the same type of model tends to standardize the way risk, opportunity, and military proportionality are perceived. Superiority is no longer determined solely by the number of strike platforms, but by the quality of the AI that configures the list of targets and the way it orders "efficiency" and "cost" of an operation.

The conflict between Anthropic and the Trump administration, between ethics and labeling as "security risk"

Anthropic's positioning regarding the military use of its models is articulated in two explicit red lines: the prohibition of use for mass internal surveillance and the refusal to engage in fully autonomous lethal weapons. These limits, formulated as a minimum ethical standard in relation to fundamental rights, come into direct collision with the expectations of the Trump administration, which sought as few constraints as possible on AI technology for a wide range of security uses.

The consequence is the transformation of a disagreement over ethical guarantees into a national security dispute. The administration decides to ban Anthropic products for federal agencies and qualifies the company as a "supply-chain risk" , a label previously reserved for foreign actors considered strategic adversaries.

From operational dependence to political exclusion

The paradox lies in the fact that this labeling follows the use of Claude in the preparation and execution of a large-scale attack on Iran . Anthropic's technology is advanced enough to be integrated into a complex strike operation, but the company is subsequently treated as a structural vulnerability and pushed out of the federal ecosystem. ​ The message that emerges is that "risk" does not only mean potential technical breaches or dependencies on external infrastructures, but also resistance to certain requirements of the military apparatus. Supplier ethics is reframed as a security issue, and the refusal to expand the uses of the model becomes a reason for political sanction.

The Pentagon shifts towards OpenAI

After the break with Anthropic, the Pentagon is turning to OpenAI, with an agreement that allows the integration of its models into classified networks for "all legally permitted purposes" . The formulation shifts the focus from material limits (what must not be done) to a formal criterion in a context where the law has not yet adapted to the capabilities of AI for massive data collection and analysis.

In practice, this opens the possibility of using AI in forms of surveillance and data analysis that most citizens would perceive as "massive," even if they fall within the letter of some outdated norms. From the state's perspective, the new provider offers maximum maneuvering space; from the perspective of democratic control, it emphasizes dependence on an insufficient legal framework to function as a true barrier.

Subsequently, OpenAI returned and modified the contract with the U.S. Department of Defense, and in an updated statement emphasizes that AI will not be used for the surveillance of citizens:

"In accordance with applicable laws, including the Fourth Amendment to the United States Constitution, the National Security Act of 1947, and the FISA Act of 1978, the AI system will not be intentionally used for the internal surveillance of individuals and American citizens.

To avoid any doubt, the Department understands that this limitation prohibits the deliberate tracking, surveillance, or monitoring of individuals or American citizens, including through obtaining or using personal or identifiable information acquired for commercial purposes."

Furthermore, Sam Altman, CEO of OpenAI, acknowledged that the company should not have rushed in announcing the recent agreement with the U.S. Department of Defense, amid negative public reactions.

"State models" and "resistance models"

This reconfiguration produces a clear fragmentation of the AI ecosystem. On one hand, there are models perceived as closely aligned with the needs and logic of the security apparatus, "state models," for which the willingness to accept broad formulations regarding use becomes a competitive advantage. On the other hand, a space of models associated with maintaining certain red lines, "resistance models," emerges, which gain reputational capital among segments of the public concerned with limiting the militarization of AI, but lose access to public contracts.

This polarization transforms the technology market into a field of political and strategic alignment. Internationally, it favors the emergence of relatively impermeable technological blocks, linked to military alliances, with divergent standards regarding surveillance, weapon autonomy, and the role of AI in lethal decisions.

Military and geopolitical implications

The case of Claude shows that military superiority is no longer determined solely by the quantity of armament or industrial capacity, but also by the degree of integration of large AI models into the operational chain. AI becomes a force multiplier: it structures, synchronizes, and prioritizes the use of existing capabilities, maximizing their effect.

In this sense, control over algorithms, their development, adjustment, and governance becomes an essential dimension of power. A state capable of combining data infrastructure, high-performing models, and political consent to use them extensively in a military context gains a structural advantage that is difficult to compensate for through conventional means.

Civil-military fusion and precedent for other powers

The use of a commercial model in a large-scale attack on Iran practically legitimizes civil-military fusion in the field of AI. What until recently was described as a risk, compelling private companies to directly support military projects, becomes a de facto standard in a context of global competition.

For other powers, the episode serves as a precedent and justification. States that are already pursuing close integration of the tech sector into the military apparatus can invoke the American example to normalize pressures on their own companies. The likely result is an acceleration of the global military AI race, in which the argument "others are already doing it" undermines any initiative to establish common red lines regarding surveillance and lethal autonomy.

Responsibility and legitimacy in AI-mediated lethal decisions

The entry of AI into the target selection process and the tactical loop fragments the classic chain of responsibility. The lethal decision is no longer the exclusive result of a commander's judgment, but of the interaction between model developers, company leadership, military configurators of the systems, and political actors who authorize integration.

Without a specific legal framework, each actor can invoke the limits of their own control. The company can argue that it provided only a general tool; the military can claim that it used a system deemed safe; the political actor can invoke the information available. From the perspective of humanitarian law, it becomes more difficult to identify decision-making responsibility for a major error or for a systemic abusive practice.

Transforming military culture

In the medium term, this distribution of responsibility risks altering military decision-making culture. If operators and commanders perceive that their decision is "confirmed" by an AI system, there may be a relaxation of internal prudence barriers and a tendency to partially externalize the moral burden onto technology.

This affects not only individual accountability mechanisms but also public perception of the legitimacy of using force. A society that knows that lethal decisions are mediated by opaque models may become more suspicious of the official narrative regarding the proportionality and necessity of operations.

From the episode of using AI in the attack on Iran to the new normal

Analyzed together, the use of Claude in the attack on Iran, the conflict between Anthropic and the Trump administration, and the rapid pivot towards OpenAI outline the beginning of a new normal, not an isolated case.

Commercial artificial intelligence is confirmed as critical infrastructure for planning and conducting war, and the status of providers is conditioned by their willingness to accept broader uses of models.

The solid hypothesis is that this episode marks the transition from the experimental phase to the phase of systematic integration of commercial AI into the military apparatus. In the next stage, the central stakes will no longer be related solely to technical performance but to the control regime, the accepted ethical and legal limits, and how democratic societies succeed or fail to maintain a clear accountability framework in the face of a technology that profoundly restructures how lethal decisions are made.

Synthesis made with the help of a data monitoring flow provided by the media monitoring platform NewsVibe Romania. The analysis presented has been enhanced with the help of Machine Learning and Artificial Intelligence tools.