The Kraken Is Reading

This week: detecting CoT hijacking, AI and the enclosure of the scientific commons, responses to AI in math, obsolete facts in agent memory, a model of geo-economic competition

Detecting Chain-of-Thought Hijacking in Backdoored Multimodal Large Reasoning Models

“A CoT hijack changes the model’s mind almost for free, and it’s hijacking tokens are image-dependent”

“Chain-of-thought (CoT) reasoning enables multimodal large reasoning models (MLRMs) to
solve complex visual problems, but also exposes them to CoT hijacking: an adversary implants a backdoor through poisoned fine-tuning so that a visual trigger induces a hijacking turn, redirecting an otherwise faithful reasoning chain to an attacker-chosen answer.”

Zhang-1.png

“Worse still, even when the reasoning trace is inspected, the hijack can remain difficult
to identify: a benign model may also initially reason consistently with the image before redirecting its answer through a turn that resembles ordinary self-correction (DeepSeek-AI, 2025; Yang et al., 2025; Gandhi et al., 2025), such as “Wait, let me think. . . ” (Dou et al., 2026). As more multimodal models (OpenAI, 2025b; Gemini Team, 2025; Bai et al., 2025a; Kimi Team, 2025; GLM-V Team, 2025) adopt CoT reasoning, detecting hijacks concealed within seemingly legitimate reasoning traces becomes increasingly important.”

Zhang-2.png

“We presented INERTIAGUARD, a detector of CoT hijacking based on visual belief inertia. By relating belief revision to entropy within imagedependent reasoning spans, it identifies suspicious
answer shifts that evade existing token-level monitors. Across four backdoors and four benchmarks, INERTIAGUARD achieves high AUROC and true positive rates at low false positive rates in both indomain and out-of-domain evaluations. Ablation studies further demonstrate the value of combining image localization, entropy, and span-level belief revision. These results highlight visual belief inertia as an effective signal for detecting reasoning level backdoor attacks against multimodal models.”

Zhang, Z., Wei, Q., Chen, W., Maennel, O., Zhang, W. E., Feng, L., & Xu, M. (2026). Detecting Chain-of-Thought Hijacking in Backdoored Multimodal Large Reasoning Models. arXiv preprint arXiv:2610.16598.

https://arxiv.org/abs/2610.16598

AI and The Enclosure of the Scientific Commons

Incredibly well written and should give pause

“This article seeks to provide a framework for understanding why some of us may be feeling a bit Luddite these days. It highlights a related but separate phenomenon from the time of English industrialization: the enclosure movement. In many ways this article is a primer on the concept of the commons, both historical and present-day, aimed at an audience of scientists and engineers who may not have been exposed to the related history and theory before.”

“…it remains true that the commons of preindustrial England no longer meaningfully exist: present-day common lands are predominatedly used for recreation and conservation, rather than supporting livelihoods (A. J. L. Winchester 2022). If we are to identify a real, singular tragedy of the historical commons, it is not that they collapsed; it is that they were stolen.”

“While commons in the traditional sense are still studied (e.g., E. Ostrom 1990), modern scholarship is often concerned with a non-traditional form: information commons. In economic terms, information is unlike the pastures of yore in that it is non-rivalrous (if I’m using it, that doesn’t prevent you from using it as well) and nonexcludable (it is expensive to prevent someone from accessing and using it). These characteristics make information resistant to being subjected to a private-property regime, and correspondingly make it amenable to accumulation in a commons.”

“Finally: in this interpretation, scientists and scholars are commoners, in the literal meaning that their way of life is bound to the commons, existing somewhat outside of the market system. Unlike agricultural commoners, of course, they generally enjoy high socio-economic status, an attribute they share with the Luddites. Indeed, many of them consider the work they do, which creates and sustains the commons, to be fulfilling and rewarding.”

“We are presently witnessing a dark version of “Moglen’s Metaphorical Corollary to Faraday’s Law,” mentioned in Section 4. It is now clear that if you give people generative AI tools and a network, they will spam slop. Virtually every commons-based peer production process mentioned above is currently facing the equivalent of a distributed denial-of-service (DDOS) attack on its integration processes.”

“Kranzberg’s First Law of Technology pithily captures a profound truth. Specific technologies do not have intrinsic moral content; but the ways that they are used in society absolutely do. It is a category error (as well as a significant generalization) to say that AI is “good” or “bad.” But the ways in which people choose, or are coerced, to use and develop it are emphatically not neutral.”

Williams, P. K. G. (2026). AI and The Enclosure of the Scientific Commons. arXiv preprint arXiv:2610.27251.

https://arxiv.org/abs/2610.27251

Early career researcher responses to AI in math

“Train their kraken”

“In September and early October, we surveyed over 260 early career mathematicians
representing nearly 60 institutions to get their thoughts on AI in math. Here, we share the
survey results.”

brandenberger-1.png

“‘I am very concerned that the standard expectation for a job will be dozens of papers now, cranked out by people that don’t fully understand them, rather than slower more thoughtful work.’”

brandenberger-2.png

“‘Math used to have 2 sides, solving and explaining. It seems like the human role is shifting to be just explaining. So specializing more in teaching and pedagogy seems increasingly important and may buy better job security.’”

“‘It’s been quite stressful to focus on work properly. Labs using my dream problems to train their kraken is certainly not pleasant. On the other hand, it forced me to reflect a bit deeper about my values and such.’”

Brandenberger, A., Mularczyk, H., & Woodruff, D. (2026). Early career researcher responses to AI in math. arXiv preprint arXiv:2610.40584.

https://arxiv.org/pdf/2610.40584

When Retrieval Is Not Enough: Obsolete Facts in Agent Memory

“…obsolete evidence can override a retrieved correction”

“Memory-agent benchmarks often collapse retrieval and answer generation into one score, making it hard to tell whether a failure comes from missing an update or from not using an update that was already retrieved. We examine this distinction on the single-hop FactConsolidation task of MemoryAgentBench.”

“Persistent memory lets an agent carry preferences, task state, and factual updates across sessions. A deployed system must therefore solve at least two distinct problems: retrieving the relevant information, and selecting the currently valid version when memory contains both an old fact and a later correction. End-to-end accuracy hides this distinction. The retriever may return the newest fact, yet the language model can still answer with the older one. We ask the following question: when the correct update is retrieved, how do evidence budget, unrelated distractors, and an older competing fact affect whether an LLM uses it?”

“These results suggest that retrieving the correct fact is not enough. An obsolete competing statement can be more harmful than an unrelated piece of evidence of the same size.”

“Within these boundaries, when both versions are present, obsolete evidence can override a retrieved correction.”

Poikarayil, O. B., Babu, T. B., Nahhas, A., Mehta, H. D., Baktash, J. A., De Luca, E. W., & Turowski, K. (2026). When Retrieval Is Not Enough: Obsolete Facts in Agent Memory. arXiv preprint arXiv:2610.36308.

https://arxiv.org/abs/2610.36308

The Great Game: A Model of Geoeconomic Competition

Spheres of influence…visualized in a sphere.

“We build a model of two hegemons that have valuable trading relationships with each other and at the same time compete in exerting geoeconomic power over countries in the rest of the world.”

Clayton-1.png

“Countries in the rest of the world benefit, all else equal, from hegemons containing each other since they partially squander their power with offsetting demands. However, containment policy can also worsen the efficiency of global production, leaving the rest of the world worse off.”

Clayton-2.png

“Hegemons trade off the economic rents they can earn from selling upstream goods to their rival, against the fact that more efficient production enables their rival to better compete in geoeconomic power over countries in the rest of the world. We show that the model gives rise to regimes of containment, in which a hegemon tries to weaken its rival through a combination of export wedges and demands that countries in the rest of the world buy less of the rival’s goods, or accommodation, in which a hegemon promotes its rival and focuses on economic profits.”

Clayton-3.png

“A hegemon is more likely to contain its rival when the goods they use to form economic threats against the rest of the world are more substitutable in the rest of world countries’ production; when the two hegemons are less interdependent with each other; and when the geopolitical rivalry between hegemons is higher. We characterize how geoeconomic rivalry affects the welfare of the rest of the world: at low levels of rivalry the rest of the world benefits due to its better bargaining position, but as rivalry increases further the rest of the world starts to suffer from the loss of economic efficiency in the world economy induced by the stronger attempt of each hegemon to contain the rival.”

Clayton-4.png

Clayton, C., Maggiori, M., & Schreger, J. (2026). The Great Game: A Model of Geoeconomic Competition. Available at SSRN 6460778.

http://www.nber.org/papers/w35952

Reader Feedback

“Some people, in particular, then tend to over-state the potential of a given technology. With the weight modification LLMs for instance, what they did in that paper, it was more like bug fixing than a fundamental improvement. That’s where it’s at right now.”

Footnotes

That transition from knowledge consumption to knowledge production is a lot of fun. It’s a hard turn. I’ve been focused on consumption to get to the frontier and signal a research agenda. That’s all signalled. I reckon that some of the right people are aware of what I’m up to.

There’s just so much literature. It seems like some of it has started reading and appreciating itself.

So, here I go making.

Never miss a single issue

Be the first to know. Subscribe now to get the gatodo newsletter delivered straight to your inbox

Subscribe to gatodo

Don’t miss out on the latest issues. Sign up now to get access to the library of members-only issues.
jamie@example.com
Subscribe