Click here to read our latest report “Beyond Extremism: Platform Responses to Online Subcultures of Nihilistic Violence”

AI Bypassing the Expertise Barrier: The Houthis, Claude Code and a New Pathway to Missile Proliferation

AI Bypassing the Expertise Barrier: The Houthis, Claude Code and a New Pathway to Missile Proliferation
16th November 2026 Akash Shah
In Insights

In its September 2026 report on ‘Detecting and countering misuse of AI’, Anthropic disclosed a weapons-development cell operating in northern Yemen, in an area under Houthi control. The cell used its Claude Code tool to build guidance, navigation and control (GNC) software for missiles. The group used the AI tool to circumvent a common shortcoming among non-state actors: a lack of trained, qualified human resources for high-end software engineering solutions. The report details how the cell ran several simultaneous instances of Claude, each assigned a distinct role, which Anthropic described as functioning “akin to how a lead would delegate work on a small engineering team” (p.113). The workflow passed incrementally through one instance writing the desired code, another conducting research, and the third reviewing the work for accuracy. According to the report, three concurrent programs were underway. These included a guided rocket that even reached the test-firing stage, a multi-stage ballistic missile with an intended range beyond 2,000 kilometres, and the multi-variant missile “R2000”, which included a hypersonic glide vehicle (p. 112). This is the first publicly documented instance, at this scale, of a non-state group replacing an engineering team with a commercial AI coding agent to improve the precision and range of ballistic missiles. 

However, the 2026 report is not an isolated case. The same report mentions the GTG-27005 case where a freelance team used Claude to design an autonomous FPV (first-person-view) drone swarm (p. 117). Although the team was based in Russia, Anthropic assessed that it was not a Russian state entity. The team used real battlefield footage from Ukraine to design a resilient targeting logic for the swarm. The fact that two unrelated non-state sectors, from two different regions and conflicts, have independently used a commercially available AI tool to compensate for a shortage of specialised human resources hints at a larger pattern. A documented trend shows tactics diffusing among different non-state actors. It is possible that more groups will follow this pattern if the gap in AI governance is not addressed soon, and the reason this case deserves sustained analytical attention. 

This Insight analyses an emerging trend of AI-assisted sophisticated weapons development programs among non-state actors and its implications for weapons proliferation and AI governance. Using the case study of the Houthis’ interaction with Claude AI, the Insight examines how commercial AI tools can overcome the scarce human-expertise barrier that has historically prevented non-state actors from pursuing high-end weapons like guided rockets. It further delves into why this analysis matters beyond Yemen and offers policy recommendations for AI developers, governments and counter-terrorism practitioners to mitigate the risk of AI misuse. 

The Pattern of AI misuse

Anthropic highlighted six AI misuse cases in its September 2026 report, spanning Yemen, Russia, and China. Alongside the Houthi-linked missile development and Russian non-state drone swarm cell, the report also cites the Claude use cases by Chinese and Iranian state actors for naval fire-control, electronic warfare and open-source targeting analysis, respectively. The capability curve is moving across multiple actors at the same time, at different levels, including militaries, state-sponsored proxies, and autonomous non-state cells. The common denominator across all actors is commercially available AI models that drastically shorten the intent-to-capability loop. 

These tactics readily diffuse across the non-state actor ecosystem, regardless of ideological inclinations. Research shows that extremist groups learn from one another, even from ideologically divergent movements, and adopt effective tactics in their own operations. The Yemen case is unlikely to remain an exception. The Houthis’ close ties to Iran suggest that other semi-autonomous groups affiliated with Iran could follow the same approach.  As Anthropic specifies in the report, the safeguard protocols blocked many of these nefarious requests, but not all. The non-state actors tried to obscure the end use of the requested information by spreading queries out and splitting the project across several sessions; Claude’s safety checks were bypassed by breaking the project into several small requests distributed across separate conversations. Claude’s built-in safeguards largely treated each request as a stand-alone piece of information to determine intent; essentially a single-query, intent-based moderation framework. This implies that a series of questions, each seemingly harmless on its own, could add up to a weapons project when put together, as happened in this case. As a result, no red flags were raised. This weakness in the underlying safeguard framework is not unique to Claude. Tech Against Terrorism’s Counter-Terrorism AI Benchmark tested 173 models and found that users can disguise themselves as a safety researcher and the model is far more likely to help with the requested information. It shows how easily safeguards, particularly those based only on intent signals, can be misled by malicious actors. Furthermore, this was not a one-off effort, and the group remained in dialogue with the AI. When the rocket tests didn’t yield the desired results, it reportedly went back to Claude to figure out what went wrong and look for ways to fix it. This highlights a method that can be repeated and adjusted to correct mistakes, while the built-in safety system, which analyses one request at a time, may struggle to flag the intent behind those requests.

Bridging the Expertise Barrier

Specialised human resources and, to a large extent, limited access to materials and fabrication facilities have been the biggest inhibiting factors in the proliferation of missiles to non-state actors. GNC is a highly specialised field which requires knowledge of signal processing and flight dynamics. Only a handful of countries can domestically produce ballistic and cruise missiles, while others must rely on imports from other states. Non-state actors have traditionally had to rely on state patronage to access high-end weaponry, such as the United States providing Stinger anti-aircraft guided missiles to Afghan Mujahideen during the Soviet invasion of Afghanistan. For a long time, this scarcity acted as a ceiling against the proliferation of guided and precise munitions to non-state actors on their own, making them rely exclusively on external help. However, non-state actors now have access to niche specialisation that compounds the risk of proliferation, courtesy of AI. 

The Yemen case is the first public instance of the ceiling beginning to lift through a channel no existing export control regime is built to monitor. Commercial software subscriptions for services like Claude are available worldwide. It is worth noting that software is just one element of advanced weapons production. Building a guided missile also requires precise manufacturing, dependable propellants, warhead integration and testing infrastructure, all of which pose significant challenges of their own.

Anthropic also explicitly stated that it has no evidence that this interaction with its AI chatbot produced a working weapon (p. 113). The episode essentially means that one of the scarcest and least substitutable inputs in this chain, specialist software engineering expertise, is no longer a barrier for non-state actors. 

Broader Implications

The Houthi movement has previously possessed missile capabilities, but they were never domestic in any meaningful sense. Multiple reports from the UN Panel of Experts on Yemen have documented the transfer of Iranian-origin short-range ballistic, cruise and anti-ship missiles to the group. This transfer violates the UN arms embargo imposed in 2015. Missile fragments recovered after the attack on Saudi Arabia had manufacturer markings of the Shahid Bakeri Industrial Group, an Iranian defence firm. Recently, shipping strategies have shifted from moving whole missiles to components like engines, guidance electronics, and rocket motors to avoid interception. These components are assembled inside Houthi-controlled territory, supported by technicians from Iran’s IRGC and other affiliated militant groups. UN experts have described the Houthis’ transformation from a lightly armed local group to a potent force capable of targeting the Red Sea shipping lane and neighbouring countries.

The network of Iran-aligned armed groups spans Lebanon’s Hezbollah, Kataib Hezbollah in Iraq and the Houthis in Yemen. Each group’s individual technical, financial and operational autonomy varies considerably; however, they share overlapping technical and training relationships. From the Yemen case, it is reasonable to infer that orchestrating parallel AI instances to substitute for an engineering team is transferable tradecraft. The lessons can easily spread through existing information–sharing ecosystems on social media, via informal channels, chat groups, and shared trainers across these groups. 

The diffusion risk is not constrained by ideological alignment between non-state actors. Conceptual knowledge that large language models can substitute for scarce engineering talent in weapons development is now public record. Extensive international media coverage of Anthropic’s disclosure has already ensured this. The expertise gap could now ostensibly be bridged, leaving non-state actors to focus on the remaining critical steps and stages for weapons development programs. Groups like the Houthis, with a history of tweaking and repurposing sophisticated military hardware, are likely to have an advantage in further developing AI-assisted weapons programs. The diffusion risk is what makes this case of AI misuse stand out from conventional proliferation hazards, as it is not primarily about moving dangerous material from one group to another, but about a relatively simple method of circumventing guardrails. 

Policy Recommendations

AI companies understandably do not disclose much about the employed safeguards; Anthropic’s report, however, did reveal their intent-based safeguard framework. The recommendations offer plausible solutions to close the gap in this particular framework. The cell in Yemen was able to bypass built-in safeguards with a single query in an intent-based content moderation framework. This is a lesson for AI developers: they need to move toward behavioural detection that can recognise intent across fragmented sessions and accounts. The problem is that, most of the time, the information users seek is dual-use. Flight control and guidance simulations are standard engineering and robotics tasks, and users could range from hobbyists to engineers working at licensed companies. To reduce overwhelming false positives, since most AI users have legitimate reasons for these queries, developers could use a layered, multi-signal detection technique that combines behavioural and semantic clustering across accounts with account-linkage signals. It could help categorise risk proportionality and, once it crosses a threshold, flag and escalate it for human review. 

Traditional export control regimes like the Missile Technology Control Regime (MTCR) and the Wassenaar Arrangement were built to govern the movement of physical material and technical documents to help prevent proliferation of missile technologies. They are insufficient to counter the software-enabled engineering assistance for weapons development, accessible through a commercial interface. Their mandate needs to be upgraded to include an explicit category for AI-assisted capability transfer, along with AI-assisted sanctions and monitoring mechanisms. While Anthropic’s voluntary disclosure is commendable, international watchdogs, like the UN Panel of Experts on Yemen in this case, should be resourced and mandated to request relevant threats directly from AI companies to foster regular, agency-level cooperation. 

For counter-terrorism and extremism practitioners, threat assessments of non-state groups should incorporate AI-assisted technical capabilities alongside traditional indicators such as financing, training, weapons supply chains, and material transfer. These reports should be treated as intelligence enhancements and integrated into existing monitoring of extremist networks’ capacity building. 

Conclusion

Anthropic explicitly states in its report that no group was able to produce or operationalise any weapon and that the cases it identifies represent the most sophisticated AI misuse it could find. Anthropic’s safeguards include aggregate monitoring of the content, but the real test is their performance against fragmented requests. While the overwhelming majority of user queries fall well short of the danger posed by these cases, even a fraction of users with evil intentions could have larger ramifications. Non-state actors, with a history of state-sponsored missile transfers, used a commercial AI coding tool to perform tasks that have historically been a state’s domain. This is more than just proof of an imminent new threat; it is a credible indicator of an existing one. Analysts have also warned that simply banning accounts is not enough; the groups could simply migrate to another platform or to open-weight models. Individual companies cannot solve this particular problem alone; hence, industry-wide cooperation is needed in this regard. The Global Internet Forum to Counter Terrorism’s response to the exploitation of social media by extremists offers a precedent that could be adapted by AI developers for sharing attempted exploitations of their models or any weapons-related misuse. The monitoring and governance architecture at multiple tiers needs to be built soon. It also needs to account for further iterations of this approach to stay one step ahead of actors misusing AI. 

–

Akash Shah is a Research Officer at Strategic Vision Institute, Islamabad. His research focuses on the security dimension of Outer Space and Emerging Technologies. 

–

Are you a tech company interested in strengthening your capacity to counter terrorist and violent extremist activity online? Apply for GIFCT membership to join over 30 other tech platforms working together to prevent terrorists and violent extremists from exploiting online platforms by leveraging technology, expertise, and cross-sector partnerships.

The views and opinions expressed in this Insight are the authors’ own and do not necessarily reflect those of GNET or any of its partners.