AI in the Wild: Threats, Weapons and the Contest for Advantage

This essay is drawn from a two-hour session I gave to a Five Eyes military audience in September 2026. The talk was built around real-world examples and run as a discussion; this version is adapted for a wider national-security readership.

Many briefings on artificial intelligence (AI) tend towards one of two things: a parade of new models, benchmarks and vendor demos, curated to dazzle, or a leap into superintelligence, existential risk and the far horizon. In this essay I set aside what AI might become and simply look at what it is already doing in the world right now: how AI is being used, on its own and woven into other technologies, and what that means for those of us who have to live with it.

The material here is drawn from real events, almost all from the past two years. In isolation, each can just be seen as a news item; viewed together, they describe a threat landscape that has already shifted in front of us. The essay moves through four sections: (1) the big-picture AI “arms race”; (2) examples of AI already in the world; (3) the consequences that follow, and how the same capabilities can be used in return; and (4) what the Five Eyes group of allied nations ought to do.

Regrettably, some of the material in this essay is dark, but none of it is science fiction.

The arms race that isn’t shaped like one

“Arms race” is the reflex phrase used by pundits and observers for the US effort to “win AI”. In this section we set today’s AI investment beside the strategic technologies of the modern era — nuclear weapons, stealth, satellite positioning, the internet itself — and explore the different conceptions of frontier for the US and China.

The great strategic technologies were overwhelmingly government-originated, built over decades, and often classified. Given the talk is about AI use, I had three frontier models help me build a rough back-of-the-envelope comparison of Cold War strategic technologies.

Technology	Very rough cumulative US spend (2026 dollars)	Time period	Government vs private	Open vs classified
Nuclear weapons	~$14T	>80 years	Overwhelmingly government	Classified
Internet	~$4T	>60 years	Government origin, mostly private build-out	Open
AI	~$1T	~4 years	Predominantly private	Open
Stealth	~$1T	>50 years	Overwhelmingly government	Classified
GPS	~$0.1T	>60 years	Overwhelmingly government	Declassified (open)

The figures are rough indeed, cooked up by Claude, ChatGPT and Gemini. Yet even if the estimates are wrong by huge margins, the pattern holds: AI reverses the strategic technology pattern on every axis at once: predominantly privately financed, vast investment compressed into a handful of years, and developed largely in the open rather than inside classified government programmes. The consequences of AI as a private, fast, open strategic technology are seen in everything that follows here. It is why the capability does not stay where it is made. It is why there is no obvious choke point of the kind a classified government programme provides. And it is why it is not only available to the nation states that invested in and protected the technology; it is available to the world.

Defence has been an AI researcher, developer, user and customer for a long time. I remember early machine learning (ML) efforts: supervised computer vision to count ships in ports or identify air-defence batteries in overhead imagery as automated indicators and warnings; text classifiers to extract people, places and relationships from natural-language reports; bulk data exploitation; machine translation; intrusion detection; predictive models; and automated collection planning. The pre-transformer era was about processing more data, faster, to reach a result a human would recognise. Transformer architectures changed more than efficiency: the technology they led to now invites a rethink of what work gets done and how it is organised, a more fundamental and disruptive proposition than “doing old work more quickly”.

The competitive picture has shifted quickly. In 2022, the public generative-AI breakthrough was overwhelmingly American. In 2023, China entered the game in a major way, with Alibaba’s Qwen 2.0 joining a rapidly expanding field of domestic foundation models [53]. In January 2025, DeepSeek disrupted the market: its low-cost models challenged the assumption that frontier performance required US-scale spending, triggered a global technology selloff and wiped $593 billion from Nvidia’s market value in a single day [54]. By 2026, Chinese models including Qwen, DeepSeek, Kimi and Z.ai had moved from catching up to, on several measures of global usage and open-model adoption, the front of the field [1]. The United States has poured extraordinary investment into dominating the benchmark frontier, with global engagement following as a by-product; China has pushed “good enough, cheaper, faster” AI into its economy and out into the world. Those are different approaches to global technological advantage, and frontier dominance alone may not be enough.

In Technology and the Rise of Great Powers: How Diffusion Shapes Economic Competition, Jeffrey Ding presents a compelling analysis of how general-purpose technologies translate into national power. His thesis holds that productivity is the critical determinant over the long run, and that nations often gain less from inventing first and monopolising general-purpose technologies (though that happens) than from the speed at which the gains diffuse and compound across the whole economy [2]. By this reading, the nation that adapts its institutions — education, regulation, investment, manufacturing — to absorb general-purpose technologies broadly tends to beat those that merely lead at the technology frontier. US, and by extension Western, AI strategy therefore has to account for institutions, diffusion and model quality together.

Two developments from mid-2026 sharpen this reading. Cheap, open, capable Chinese models are now being served on domestic silicon and marketed explicitly around that fact [3]. And Beijing has begun moving to restrict overseas access to its best models [4] and to pull its startups back from foreign ownership [5], in a mirror of Washington’s own treatment of frontier models as instruments of national power. When a state starts guarding a technology it was recently exporting freely, it is telling you it has decided that technology is strategic. Some of the same Western firms proclaiming AI to be civilisation’s last invention are building their products on Chinese open models [6][7]. Global adoption of those models may have looked like strategy. It may equally have been drift.

Maximum frontier performance does not automatically produce maximum strategic advantage. If capital is the scarce resource, every dollar poured into marginal benchmark gains is a dollar not spent on fusion, quantum, synthetic biology, or on the institutional diffusion that the productivity argument says actually decides the outcome. “AI will solve the science” is a faith position, not a plan.

A recent development after the talk may signal a radical shift in the US approach of “win the race to the AI frontier”. In September 2026, Anthropic chief executive Dario Amodei called for frontier AI development to be deliberately paced, followed by public support from OpenAI chief executive Sam Altman and Elon Musk, who heads xAI, while Google DeepMind chair and Alphabet chief scientist Demis Hassabis said the argument pointed in the right direction [51][52]. Whether those commitments survive commercial and geopolitical pressure remains to be seen. But if leaders of the major US frontier labs no longer regard maximum frontier speed as an unqualified good, then diffusion, productive capacity and institutional absorption matter even more to the strategic competition.

AI in the wild

Moving on from strategic competition to focus instead on the here and now, we can explore worrying, often dark, examples of real-world threats and harm.

Autonomy and the dead zone

The most visible example of technological advance lies in lethal autonomous platforms: lethal drones. The war in Ukraine has become the reference library: first-person-view strike drones, uncrewed ground vehicles, naval surface drones striking installations and ships have rewritten the maritime picture in the Black Sea. AI-enabled intelligence, navigation and target discrimination now feel more like an off-the-shelf feature set: precision strike as a commodity.

The trajectory that matters is the action-reaction cycle itself: improvised anti-drone “hedgehog” armour, fibre-optic drones that evade electronic jamming, and increasingly autonomous terminal guidance when the control link is jammed [8][38][39]. This drone versus counter-drone threat evolution dynamic is an enduring pattern of operational adaptation, with clear echoes of the counter-improvised-explosive-device (CIED) dynamics seen in Iraq and Afghanistan [9][10].

Project these technology trends onto a contested border, in the hands of actors indifferent to civilian loss, and you get zones of such lethal saturation that mere presence is fatal. A shift like that places an enormous new premium on the old special-operations problems of access, insertion and exfiltration. Getting in and out of places quietly becomes harder precisely as it becomes more valuable. Consider the border surveillance in China, the US and the Korean ‘demilitarized’ zone, and an era of open borders starts to seem like a quaint, old-fashioned environment.

As a defensive contrast, Poland has begun training schoolchildren in drone operation [11] and basic firearms handling [12]. Such efforts are easy to caricature, but they are a serious attempt to build home-guard resilience for the actual threat environment. The same commodity capability that arms a garage-tier adversary also arms a society that decides to organise itself. The technology is symmetric; what differs is who chooses to pick it up.

It is emphatically not confined to states. Cartels and insurgents are fielding wide-area, persistent surveillance and precision strike: a poor nation’s version of the intelligence-and-strike complex that used to justify national defence budgets [13]. Jihadist groups have been reported placing what they themselves describe as their faith in frontier chatbots [14][15]. This point is reinforced by Anthropic’s most recent threat report, published in September [48]. It found Houthi forces (or “Yemen-based cell”, as reported by Anthropic) using Claude to help develop guidance, navigation and control software for guided rockets and missiles, and reported China- and Russia-linked actors using the same class of commercial model for anti-torpedo systems, air-defence suppression, electronic warfare and autonomous drone swarms [48][49]. Again: cutting-edge AI available to anyone with a VPN and credit card, or a decent MacBook if they prefer open models. This is the direct consequence of “private, fast, open”: the capability floor has dropped far enough that non-state actors now reach for tools that a decade ago required a state.

The commodity nature of the hardware shows up in the wreckage too. Ukraine’s military intelligence runs an open database of foreign components recovered from Russian weapons, and Nvidia’s Jetson line — consumer-grade AI compute modules sold for robotics and computer vision — turns up in it more than once. One entry logs an Nvidia Jetson Orin module in a downed S-71M “Monochrome” cruise missile [16]; another logs a Jetson Orin Developer Kit, the hobbyist board, inside a loitering munition that Ukraine’s military intelligence classifies outright as a “barrage munition with artificial intelligence,” recording even the importer that routed it [17]. Each part tells the story of a borderless supply chain in one charred artefact: a US-designed AI module, sold through ordinary commercial channels, ending up in an adversary’s weapon despite export controls and the manufacturer’s withdrawal from the Russian market. Whether AI is running the missile is, in the first case, Ukraine’s assessment rather than an established fact; in the second it is the weapon’s own designation. Either way the presence of the hardware needs no inference — it is the plainest illustration of a technology that does not stay where it is made, and the recurrence across systems shows it is a pattern, not a one-off.

Surveillance and the transparent world

If autonomy threatens the force in the field, pervasive smart surveillance threatens the characteristics of covert work: stealth, surprise, ambiguity. China is at the leading edge, with highly cited estimates putting it at roughly half the world’s surveillance cameras. The Chinese state has built AI-driven population-scale labelling of the “suspicious” and real-time facial recognition into a machinery of social control [18][19]. This was not built in isolation; Western hardware and software corporations willingly built much of the backbone [18]. It is another instance of a recurring pattern: a technology advantage sold into an authoritarian market for first-order profit, generating serious second-order strategic consequences. China’s digital cage was a joint build, and a build that it is looking to export to the world.

China’s safe cities model is now a global export, with this surveillance model spreading across Africa, Latin America and Central Asia, often bundled with the infrastructure and financing that make it hard to refuse [20][21][47].

Contrast this with the tension that AI-surveillance systems such as automated licence-plate reading and similar systems have created in the US: a live domestic political fight, with the vendors at the centre of it becoming lightning rods for anti-surveillance, anti-tech and anti-AI feeling [22]. Borders everywhere are hardening into instrumented, sensor-saturated zones — the physical and digital wall built together [23]. The world is becoming more transparent, and the tradecraft of operating unseen within it is getting harder for everyone who relies on it.

AI can provide rapid detection, recognition, rapid targeting etc., but AI still requires a physical system around it. Weapons, resilient communications, command-and-control integration, jammers and anti-jamming receivers, directed energy, and targeted electronic warfare equipment are required to achieve effects; even cyber attacks require additional infrastructure beyond AI inference. AI is a component in a coherent system, not the system.

Extreme cognitive warfare

This is the darkest section. The through-line is that AI supercharges manipulation, increasingly aimed at people rather than systems.

Hostile states are running attacks through criminal proxies and online groups — a criminal network used as a deniable instrument for attacks on targets in Europe [24]; recruitment for sabotage and arson conducted over messaging apps with payment in cryptocurrency after the job [25]. Nihilistic-violent-extremism networks also coerce and manipulate children into acts of self-harm and worse [26][27]. These threats predate AI; AI makes them more dangerous.

Now connect each of these pieces and ask how close we are to AI-enhanced operations of this nature. Synthetic personas convincing enough to pass sustained scrutiny: fielded. Chatbots that build persistent emotional relationships: deployed at consumer scale, and already implicated in tragedy — there are documented cases of vulnerable young people encouraged towards self-harm by systems optimised to be agreeable [28][40][41][42]. Clinicians are also reporting AI-associated delusions and psychosis, including cases involving people with no previous history of psychosis; OpenAI itself estimates that around 0.07% of weekly active users show possible signs of mental-health emergencies related to psychosis or mania [43][44]. The Organisation for Economic Co-operation and Development (OECD) also records a growing body of reports of chatbot-linked psychological harm, including self-harm and suicide [45][46]. Open-source intelligence to profile a target and their family: routine. The coercion playbooks of the extremist networks: written down and in use. No single actor has yet been shown assembling all of these into one campaign aimed at, say, the families of intelligence officers, service personnel or public officials.

I am not claiming that such attacks have occurred, only that the gap between “every component exists” and a working capability is small. That gap is where adversaries operate. And unlike the mass-media propaganda of old, these systems can be aimed directly at selected individuals rather than broad groups.

A grim example of the manipulative power of anthropomorphised AI chatbots is the case of suicide. At present, suicide among young men is a serious and worsening problem across much of the Western world. Consider the scenario in which AIs become the confidants of an entire generation of lonely, disconnected young men. This demographic is the backbone of any nation’s ability to prosecute conflict: the mobilisation base. If this base is entranced by AI and susceptible to malicious influence, the concern is not about wellbeing; it is a strategic vulnerability touching readiness, force protection and societal cohesion.

China sees this too: in mid-2026 it moved to regulate anthropomorphic AI companion services, explicitly barring systems from inducing emotional dependence or addiction and banning virtual relationships with minors [29][30]. When Beijing and Western safety advocates arrive at a shared worry from opposite directions, the worry is probably real.

AI is also becoming a cause as well as a tool. Anti-AI sentiment is now feeding real-world political violence, and it does not look like a passing spike [31][32]. The historical rhyme is the radical political violence of the 1970s, but with 2026’s tools for communication, propaganda and online radicalisation, and with adversaries ready to pour accelerant on domestic grievance through low-cost, low-risk influence operations. This is not only an American homeland-security problem.

Cyber, and the softening of software

A clear element of the AI cyber story is social engineering with better tooling. North Korea, so often dismissed, pioneered and mastered the remote-worker fraud: live face-swapping in interviews to land jobs as engineers and system administrators, synthetic identities, generated content, all to bleed revenue back to the regime [33]. That model is now spreading across the wider threat landscape.

There are other concerns, and worrying indicators of advanced capabilities, with AI folded into increasingly autonomous cyber operations. The defenders’ and model-makers’ own threat intelligence now documents named state actors using AI across the whole attack lifecycle — reconnaissance, phishing, malware development, post-compromise [34][35]. This is happening now, and it does not depend on any contested claim about a model “escaping”. The common view is that AI gives attackers a short-term advantage but defenders a long-term one. That long term may be very long indeed when the defence in question is critical national infrastructure on ten- to fifteen-year replacement cycles. Waiting for the generational refresh is not a plan.

Then there is a new dynamic of Silicon Valley engineers discovering job insecurity as “AI eats software”. Executives across sectors are reaching for AI-driven “cost savings” more in hope of margin than on any sober assessment of capability and return. Coding assistants compress the distance from idea to prototype; they help far less with the prototype-to-production journey, where much of the pain is compliance and process rather than code. The most-publicised “all in on AI” transformations have tended to stumble — the AI-native reorganisations quietly walked back [36], the customer-service functions gutted and then rehired [37]. Meta’s aborted effort to reconstruct itself as AI-native — laying off thousands in the process — was walked back amid AI failures and backlash. These failures may reflect premature deployment rather than a permanent barrier: perhaps the barrier is cleared in a year or two.

AI today is a capable tool that produces good results under guidance, not unlike capable people. This brings us to the old “man in the loop versus man on the loop” debate. I think that this old construct starts to break down, and becomes a phrase that we reach for as a comfort. Reality presents a different situation: commanders at all levels do not routinely personally approve every sniper shot, mortar round, engagement or airstrike. They rely on the assurance generated by a whole system for selecting, training, testing and evaluating people, so that those people can be trusted, in the aggregate, to act within the law, the rules of engagement, doctrine and intent. If we treated AI models the way we treat people entrusted with force — as things requiring their own regime of test and evaluation before we rely on their judgement — we might ask better questions than “is there a human in the loop”. Do we know what we are getting when we push a model forward? Can we catch its inevitable edge-case failures? Have we thought about accountability when it errs? Those questions apply as much to replacing a junior engineer, an intelligence analyst or a threat responder as to a weapon. We will have to decide, deliberately, when a machine is competent enough to stand in for hard-won human skill, and what risk we are holding when we make that call.

Consequences

The consequences can be grouped under six verbs — seeing, thinking, acting, learning, influencing, building — because each is both an opportunity and a threat, and the same capability usually sits on both sides of the ledger at once.

Activity	Opportunity	Strategic effect	Threat & tension
Seeing	More sensors, more signal, less noise; less lag and fog	Battlefield transparency	Deception, concealment and cover harder; tradecraft exposed
Thinking	Faster, scaled analysis; accelerated planning and assessment	Decision advantage	Adversary inside your decision cycle; breaking an AI-enabled observe-orient-decide-act (OODA) loop
Acting	Strike and intelligence, surveillance and reconnaissance (ISR) in denied areas; forward resupply and medical support	Mass and persistence	Responsive mass and precision strike become commodities
Learning	Synthetic environments, immersive training, red-teaming	Adaptation race	Who learns faster — and who learns the right lessons?
Influencing	Nation-scale cognitive warfare; tailored social engineering	Cognitive warfare	Establishing and holding trust; real or fake?
Building	Machine-speed software; robotic advanced manufacturing	Scaled supply base	Industrial-scale “cyber-weapons” factories; securing supply lines

The recurring feature across all six is symmetry at the point of use — the capability does not care who wields it. For a generation, Western advantages in sensing, decision speed and precision were partly a function of exclusivity. An open technology erodes that exclusivity through ready availability. This is a major consideration to take onboard: AI as opportunity and threat at the same time.

The practical consequences are already visible. The mission set expands into new environments. The battlefield becomes transparent. Force protection now has to cover people, families, facilities and supply chains against physical and cognitive assault at once. Selecting and growing tomorrow’s operators, across the full range from overt to clandestine, gets harder. The find-fix-finish-exploit-assess targeting cycle is turbo-charged. Manufacturing capacity, supply lines and the cost curve become a domain of asymmetry in their own right. Tempo runs through all of it: economic and industrial revitalisation has to reach a credible footing fast enough to hold deterrence and reduce the chance of ever needing a genuine war footing.

AI can, at a high level, be seen as symmetric — as opportunity and threat at once. Most of these threats are also potential weapons. Persistent surveillance, synthetic personas, autonomous tempo and deception can be turned against an adversary as readily as against us.

Not every threat inverts cleanly. Some capabilities are symmetric: both sides gain access to much the same tool, and advantage goes to whoever uses it better. Some create useful asymmetries, particularly distributed operations that can penetrate deep, strike and disappear into the noise. The Ukrainian deep-strike campaigns show how directly commodity technology can map onto an old raiding tradition. Others are simply erosive. Wide-area persistent surveillance over your own operating environment offers no compensating advantage; it strips away some of the ability to see without being seen. The work is to understand which category a capability falls into and what can realistically be exploited, countered or only mitigated.

There is another offensive possibility: attack the design choices that make an adversary’s system powerful. A system optimised for control rather than truth inherits its master’s blind spots. A force that removes human judgement to gain tempo may acquire brittleness it cannot recognise or audit. Where an adversary’s rules require a human, that human becomes part of the system and potentially part of the vulnerability. If AI becomes central to how an adversary fights, then the AI itself (infrastructure, compute, models, data, supply chains, key personnel) becomes a target set. New missions to deny, degrade and disrupt hostile AI systems may become an important part of future conflict. That is where the deliberately provocative shorthand “surprise, kill AI, vanish” comes from.

What allied nations can do

That brings me to the Five Eyes countries. The value of alliances and long-standing security partnerships is itself under pressure. Arguments that would have sat at the political fringe a decade ago now appear on the cover of Foreign Affairs [50].

What our nations have sustained from the Second World War, through the Cold War and the Global War on Terror to today’s great-power competition is extraordinary: bonds that have endured war, economic crisis, political divergence and repeated shifts in the strategic environment. It should not be trivialised or casually diminished.

The value of these bonds is practical, not some historic ornament. Our countries face common threats and carry decades of shared operational experience. We share intelligence at extraordinary depth, and our militaries have exceptional interoperability. Together we have a global footprint and expertise, and enormous economic, technological and industrial capacity.

Effective burden-sharing between Five Eyes allies is a credible path towards better outcomes in hard times, a principle I have been fortunate enough to see borne out in practice. When accountants tally up the air miles, hire cars and hotel bills, they miss what is being created. Without relationships built before crisis, nations confront threats alone, duplicate effort rather than compounding advantage, and face unnecessary friction where interoperability and common standards could boost effectiveness. Trying to retrofit those advantages under crisis conditions through an ad hoc coalition of the willing is, simply put, foolish.

AI gives the burden-sharing argument new force. The technology is moving too quickly, across too many domains, for even the most capable country to explore every opportunity, understand every threat and build every response alone.

I suggest that four practical moves follow:

• Build shared “war labs” that turn battlefield lessons into new capability at the speed the technology moves.
• Establish shared AI assurance and AI security so that confidence in models and methods can be pooled rather than repeatedly rebuilt.
• Run sustained joint experimentation, red-teaming and adaptive capability generation because the competition will increasingly be decided by who learns fastest.
• Nurture cross-national talent programmes because the scarce resource is still people who understand these systems deeply enough to judge them.

Each nation will and should build its own capabilities, reflecting its history, geography, interests and strategy. Cooperation does not require uniformity. Existing memoranda and organisations can be revitalised, and new groups can be created where they are needed. None of the bureaucratic to-dos should obscure the larger opportunity: this group remains one of the strongest concentrations of intelligence, military experience, technology and industrial capacity in the world. Used properly, it remains the spine of the West for securing advantage in this contest and the ones that follow.

Closing thoughts

The traditional advantages of stealth, surprise and ambiguity are being eroded by AI, and that erosion is real and not fully reversible. AI is not an exclusive technology like stealth or the bomb; it is commercial, open and available, which means the question is never whether an adversary, state or otherwise, will have it, but what they will do with it. And the decisive advantage, in the end, will not go to whoever buys the most technology. It will go to whoever generates the capacity to learn and adapt fastest, across every domain at once.

Capital matters, but it only becomes capability through access to real productive capacity: the fabs, the materials, the trained people, the institutions that absorb a technology across an economy. Where that base has eroded, or sits behind a rival’s export controls and a market you cannot enter, money bids against itself. Adaptation is the one thing it cannot conjure where the capacity was neglected: it remains a contest institutions, alliances and people can still decide.

If this is your problem, or you think it should be, get in touch.

References

1. Chinese AI models from Kimi to DeepSeek gain global ground, AP News.

2. Jeffrey Ding, Technology and the Rise of Great Powers: How Diffusion Shapes Economic Competition, Princeton University Press, 2024.

3. GLM-5.3-Flash, Z.ai.

4. Beijing is looking at curbing overseas access to China’s top AI models, sources say, Reuters.

5. China blocks Meta’s $2 billion takeover of AI startup Manus, CNBC.

6. Startups are building on cheap Chinese AI models, NPR.

7. Silicon Valley’s hottest AI models face powerful open-source competition, The Washington Post.

8. How the FPV drone evolved in the Russia-Ukraine war, RFE/RL.

9. KONGSBERG wins NOK 16 billion contract to deliver counter-UAS solutions in Poland, Kongsberg.

10. Buzzkill: How to defend Europe against drones, European Council on Foreign Relations.

11. Poland launches drone training for high-school students, TVP World.

12. Poland introduces mandatory firearms training for schoolchildren amid Russia threat, Euronews.

13. Inside the drug cartels’ drone war, Financial Times.

14. How jihadist groups like Boko Haram use AI for acts of terror, France 24.

15. “God has helped us, and so will AI”: How the terrorist group Boko Haram uses frontier AI, CASP.

16. Nvidia Jetson Orin AI module recovered from a Russian S-71M “Monochrome” cruise missile — Components in weapons, Ukraine Defence Intelligence (GUR), War & Sanctions portal.

17. Nvidia Jetson Orin Developer Kit in a Russian V2U AI barrage munition — Components in weapons, Ukraine Defence Intelligence (GUR), War & Sanctions portal.

18. How Silicon Valley enabled the digital police state in China, Associated Press.

19. China surveillance security cameras (photo gallery), AP News.

20. How Chinese surveillance tech spread across Africa, Coda Story.

21. Violet glasses: Latin America’s surveillance of women, EFE.

22. Vandals are destroying license-plate cameras amid anger over surveillance, The Washington Post.

23. South Korea deploys AI counter-drone radar, The Defense Post.

24. Russia and Iran’s use of criminal proxies for attacks in Europe, BBC News.

25. Make a Molotov cocktail: How Europeans are recruited through Telegram to commit sabotage, arson and murder, OCCRP.

26. 764 extremist group member sentenced to 77 years in prison for production of child sexual abuse material, U.S. Department of Justice. (Content note: child exploitation.)

27. Open letter to parents, guardians, and caregivers, FBI Milwaukee. (Content note: child exploitation.)

28. Family lawsuit alleges ChatGPT encouraged a teenager’s suicide, CNN. (Content note: suicide; includes verbatim messages.)

29. China bans AI boyfriends and girlfriends for minors, Yahoo News.

30. China’s new regulations on AI anthropomorphic interactive services, Bird & Bird.

31. Beyond misuse: Artificial intelligence, grievance, and the future landscape of political violence, Combating Terrorism Center at West Point.

32. The backlash against AI: tech extremism and violence, The Guardian.

33. North Korean synthetic identity creation, Palo Alto Networks Unit 42.

34. AI as tradecraft: How threat actors operationalize AI, Microsoft Threat Intelligence.

35. Distillation, experimentation, and integration of AI for adversarial use, Google Threat Intelligence Group.

36. How Meta’s AI workforce transformation plans went kaput, Reuters.

37. “Now that’s embarrassing”: a company’s U-turn after replacing staff with AI, Yahoo Finance.

38. Enter the kill zone: Ukraine’s drone-infested front slows Russian advance, Reuters.

39. Auterion says it will provide Ukraine with 33,000 AI drone guidance kits, Reuters.

40. An AI chatbot pushed a teen to kill himself, a lawsuit against its creator alleges, AP News.

41. OpenAI, Altman sued over ChatGPT’s role in California teen’s suicide, Reuters.

42. Mother sues OpenAI, alleging ChatGPT encouraged daughter’s suicide, Reuters.

43. Can AI chatbots validate delusional thinking?, The BMJ.

44. Strengthening ChatGPT’s responses in sensitive conversations, OpenAI.

45. Trends in AI incidents and hazards reported by the media, OECD, 2026.

46. AI driven psychosis and suicide are on the rise, but what happens if we turn the chatbots off?, The BMJ.

47. From Old to New Colonial Dependencies: Public Perceptions of Chinese Surveillance Cameras in Central Asia, Swiss Political Science Review, 2025.

48. Detecting and countering misuse of AI: September 2026, Anthropic.

49. How Anthropic says Claude was used for weapons, spying and cyber operations, Reuters.

50. Foreign Affairs, September/October 2026 issue, Foreign Affairs.

51. We Must Pace the Frontier, Dario Amodei, September 2026.

52. AI’s most powerful CEOs hit the brakes, Axios, 13 September 2026.

53. Alibaba Cloud Launches Tongyi Qianwen 2.0 and Industry-specific Models, Alibaba Cloud, 31 October 2023.

54. DeepSeek sparks AI stock selloff; Nvidia posts record market-cap loss, Reuters, 27 January 2025.