Books I've read and reviewed. Click any cover to read my full review. I also include some longer papers.
KB swings and hits again. No major updates beyond TB II, but makes some minor updates that I'm guessing will matter. I run Operator with Green (I know, I know) so I'm really only affected by the Operator changes.
This training program will make you a beast. It will make you fit. It will do it in a time-efficient manner. It will do it without much thinking involved on your part. Buy it. Read it cover to cover. Think about what you want out of it. Follow the Activation section. Revel in your easy-won gains that you've been missing for the past few years.
Boko Haram knows no favorite LLM and is above the Claude vs. GPT debates that plague nerdy circles; they use whatever gets the job done:
have used multiple leading AI systems, including ChatGPT, Claude, Gemini, Grok, Meta AI, and DeepSeek, for every stage of military activity, from mission preparation through execution to post-mission reviewI'm a bit surprised they don't use more of the base models or open-source since they are likely easier to jailbreak, albeit at the cost of quality and sophistication of responses. I guess if you can jailbreak the SOTA models you'll choose to use those.
The use cases of "weapons troubleshooting and design of explosive devices, tactical and strategic planning, operational security, and logistics" seem obvious in retrospect, but I never really thought of how terrorists would use this.
It's almost like BH has some forward-deployed engineers helping out:
They supplied laptops with VPNs and encryption software, set up accounts and managed paid subscriptions, and advised daily on prompting techniques and bypassing platform restrictions.
The author discusses analysis of online materials:
The analysis shows that AI has mostly been used to generate and translate jihadist propaganda by unofficial jihadist media groups and individual content creatorsbut explains that it's hard to get deeper without talking to the people directly. BH isn't going around posting their jailbreak prompts on the internet and bragging about how they're Fablemaxxing! (This reminds me of how much tacit knowledge is tied up in offline or members-only spaces that's not accessible to the wider internet, e.g., Discord servers. There's so much knowledge transfer going on in the world at any given moment that's effectively impossible to track.)
Also in hindsight, it makes total sense that they were quick and avid adopters of AI given ISIS's Amaq News Agency, which has extremely high-production-value videos. The resolution is crystal clear, the editing superb, etc. They were definitely using some high-ish tech editing software and keeping up with the times based on the "theme" of the video I saw (it was very anime-like and evidently meant to appeal to younger folks because who doesn't like anime????). But base rates should also be considered: ChatGPT was the fastest-growing product of all-time, meaning some terrorist dude was certainly in the mix of trying it. Terrorists can be smart! And plugged in! Do not underestimate them! Additionally, apparently BH is pretty techy:
Boko Haram has been characterized as among the most adaptable violent non-state actors, with a high capacity for innovation and rapid responses to changing operational environments ... adapted commercially available technologies for military purposes, ranging from establishing satellite internet access in remote areas of the Lake Chad Basin to weaponizing drones
I found some of their objections to gen AI kind of funny:
For example, the presenter’s face inan AI-generated news bulletin was blurred in response to criticism that showing animated human faces was forbidden (haram), though concerns persisted about depicting animated beings at all. Text-to-speech tools have generated additional controversy, with users arguing that they should not be used because they are based on the voices of “infidels” and mispronounce Arabic words commonly used in jihadist messaging.I mean, at least they're kind of consistent in abiding by the Quran, but it's humorous (?) that they're brutally murdering and terrorizing people while complaining about animated faces.
I was curious how she got the contacts, to which she answers succinctly:
High defection rates have created opportunities to interview former commanders with recent operational knowledge, while relative stability in certain urban locations makes field work in an active conflict zone possible. Importantly, my previous doctoral research in northeast Nigeria with former Boko Haram members provided the groundwork for this study. It enabled me to draw on established contacts and networks, as well as contextual understanding of the conflict and security dynamics essential for safely accessing this population.
It becomes very clear that they have an excellent hierarchy and command structure around who's to use AI and who's to just relay their requests up the chain:
a subset of participants, predominantly women in lower-ranking leadership roles (amiras), lower-ranking male commanders (amirs), and rank-and-file combatants, tended not to have knowledge of the group’s AI use.This, of course, makes sense. That person may be ineffective in using AI, could attract the wrong attention (but unlikely), and could be a liability were they captured. Consolidating the power users creates more risk, but experts are experts—let them cook and the results will come. They organized said groups with top dogs from individual specialties:
Similarly, within JAS, “people from different units form the AI unit, like a bombmaker, a computer specialist, a gun specialist, and an engineer. The top leaders form the key unit, but then there are smaller ones in different locations. We have four fiya [senior commanders] and each fiya has this unit under them.""This has proven to be a generally successful approach:
Having a separate, specialized research and development unit has been shown to be an important factor for successful terrorist innovation (Ackerman 2016).Also, Nigerians aren't the experts, instead choosing to bus in "white guys":
“The white guys came and taught us,” a former ISWAP munzir (mid-ranking commander) said,further clarifying that he was referring to operatives from Libya, France, and Arab countries with a lighter complexion than their own. “The real source is ISIS,” he added, while a former naqib (a rank below munzir) similarly stated that “[t]he key players are Arabs who are in charge of all of this AI."
These special AI units have three main responsibilities:
They are in charge of the operational AI infrastructure, managing devices and accounts ... responsible for building capacity through internal training, passing on what they learned from the foreign operatives to the military command. ... to drive the groups’ AI use by generating insights that feed directly into daily operations and strategic planning.
These guys are willing to drop cash on AI given their premium subscriptions across most, if not all, major models, which I'd argue is the best indicator (#revealedpreference) for how much someone believes in AI's capabilities. The author states this directly:
It also indicates that leadership places signifi cant value on AI, considering that this staffing decision carries an opportunity cost, since those qualified individuals are no longer contributing as fully to other areas.
Going back to the hierarchy:
“We are not allowed to access the computers,” said yet another ISWAP naqib, “but we present the problem to them to ask the AI. I don’t think there is a problem that these people cannot overcome. They are the masters. They do the analysis with the AI and give us the strategies to implement. This support extends beyond the camp. While AI unit members typically remain at base, they provide assistance to commanders in the fi eld remotely. For example, fighters relay problems they encounter, and the unit queries a model and communicates its guidance back. Such coordination is possible through ISWAP’s documented communications infrastructure, including specific vehicles that ISWAP reportedly outfi ts with satellite internet to enable communication and data sharing on the move (Samuel 2023). In one account, a respondent even described wearing a chest camera that transmitted footage back to camp, where a commander followedthe feed, “uploaded the pictures to ChatGPT to analyze the situation,” and relayed tactical adjustments back to the field. On larger engagements or operations that require specifi c expertise, designated personnel join operations in person with laptops to provide direct support.
Their actual applications are straightforward and obvious:
We mostly used it in three ways: the fi rst one is to learn how to assemble and use guns and how to manufacture bombs. The second one is for surveillance, like how to improve our surveillance strategies to monitor what is happening in our camps and also to better understand our enemy and prepare attacks. The third one is to make plans, like, when we come up with new ideas on how to attack, we ask it for tactics on how to make it work in practice to be successful.
AI also assists with troubleshooting during combat. “It happens all the time during attacks that the guns are jammed and the trigger gets stuck,” explained the former ISWAP munzir, adding that AI provided both immediate technical fi xes by teaching “how to uncouple the gun by washing it with diesel” and tactical guidance, in terms of “how to change the military formation so that fi ghters with jammed guns move to the back and others take their positions until the problem is solved.""
For example, it advised on payload weight reduction to stay within the drone’s lift capacity, and on the design of explosive release mechanisms
Quote of the year goes to:
We saw in a movie how motorcycles can jump over bridges. We used AI to learn how to do this. We gave it information, like what motorcycles we use and the distance we need to jump and so on and it gave us steps on what we have to do. We practiced a lot and kept asking questions. We dug holes and filled them with broken glass and fire to practice. 18 of us died in the process. Eight of us managed to do it. The next time we attacked, we could jump.
My rough probability of a helped-greatly-by-LLM major terrorist attack by EOY 2027 is 25%, where major is based on vibes. I don't think the good guys have had their abilities boosted that much since they already had the best gear and software and people, whereas the bad guys have gotten a much bigger boost, closing the gap and potentially allowing them to surpass roadblocks they previously couldn't:
the risk is that AI serves as a force multiplier, lowering barriers to entry for potential attackers, enhancing the capabilities of threat actors, and enabling novel attack vectors that were not previously possible, thereby expanding existing threats or introducing new ones.