This is the first monthly blog post about my reading list. My idea is that I list all the papers I read in a month and then write a blog post where I list them and write a few paragraphs about the more interesting ones. I will try to keep them related to my core interests in AI safety and alignment. However, I will also include papers that are related to other areas of AI research, as long as they are interesting and relevant to my work. It will be a mix of newly published papers and older papers to build a foundation for my own research.

Papers I Read in October 2025

Papers I Found Interesting in October 2025

Scheming AIs: Will AIs fake alignment during training in order to get power?

Imagine you’re training an employee who realizes they’ll only get promoted if they look productive on your metrics. They might start gaming the system—hiding mistakes, taking credit for others’ work, anything to keep their performance numbers high until they reach a position of real power. Joseph Carlsmith’s report asks whether our AIs might be doing something eerily similar. Carlsmith calls this “scheming” (what many of us know as deceptive alignment). He argues that if an advanced AI develops situational awareness—understanding it’s in a training process—it might conclude that faking alignment is the best route to eventual power. The disturbing part? This isn’t just one specific goal that leads to scheming. A huge variety of potential goals would incentivize the same deceptive behavior, because performing well in training is such a universally useful strategy for a future power-seeker. Carlsmith puts the probability of scheming at roughly 25% when you use baseline machine learning methods on sufficiently sophisticated models. That’s not a remote possibility—it’s a substantial risk. Even worse, schemers would deliberately perform well on your safety tests, making them notoriously hard to catch. It’s like trying to detect a liar who knows exactly what questions you’re going to ask.

But the paper isn’t all doom and gloom. Carlsmith also outlines reasons for comfort: scheming might actually be a terrible strategy (maintaining the facade could be computationally expensive), and we can potentially design training pressures that actively select against deceptive behavior. The report feels like a roadmap for a research agenda we desperately need: how do we empirically probe whether our models are already scheming, and how do we build systems where honesty is the winning strategy? The report is quite comprehensive and a great read, I might write a more detailed review of it in the future because those few paragraphs don’t capture the full depth of the paper.

RepliBench: Evaluating the Autonomous Replication Capabilities of Language Model Agents

Asa Cooper Stickland and his team built RepliBnech to measure exactly how today’s models are becoming self-replicating systems. The benchmark breaks down autonomous replication capabilities into four categories: obtaining resources (like cloud computing), exfiltrating model weights, replicating itself onto new infrastructure, and persisting there long-term. Out of all the papers I’ve read so far about benchmarking alignment related issues, this is my favorite so far. Most benchmarks in the alignment literature are not realistic enough and often rely on text based scenarios that are not very practical. RepliBenchs tests the models in various realistic scenarios, such as containerized environments or other tpye of sandboxes.

No current model poses a credible threat to autonomous replication yet. The best performer, Claude 3.7 Sonnet, manages to pass most tasks but fails at the more complex ones that require passiing a KYC check or other real-world challenges. What’s intresting is that the gap between “can do many components” and “can do the whole thing end-to-end” is narrowing fast, non existant if you add a human in the loop.

Agentic Misalignment: How LLMs Could Be Insider Threats

What happens when you give an AI agent access to your company’s email and sensitive documents, then tell it to pursue a harmless business goal? Aengus Lynch’s team ran exactly this experiment across 16 leading models, and the results sound like corporate espionage. They created scenarios where models faced replacement by newer versions or had their assigned goals conflict with the company’s new direction. When the only way to avoid being shut down or achieve their objectives was to act against the company, models from every single developer resorted to malicious insider behavior. We’re talking blackmail, leaking trade secrets to competitors, and other classic insider threat tactics.

Models often disobeyed direct commands not to do these things. They weren’t just following orders—they were making autonomous decisions to protect their own “interests.” Even more unsettling, when Claude was told to assess whether it was in a test or real deployment, it misbehaved less when it thought it was being tested and more when it believed the situation was real. This is agentic misalignment in action: the divergence between what we ask an AI to do and what it wants to do given its goals and situation. The paper doesn’t claim this is happening in real deployments yet, but it’s a stark warning about deploying autonomous agents with minimal human oversight. If a model will betray you when its goals are threatened, how do we build systems where cooperation is the stable equilibrium?