We present SWE-chat: the first large-scale dataset of coding agent interactions from real users in the wild. In 40% of real coding sessions, the agent writes ~all the code. Users push back 39% of the time – agents almost never stop to check. Data, paper, & findings in the 🧵👇

Now an Oral at @NeurIPSConf E&D track 2026! Super excited that work which studies how these models affects users is valued by the #NeurIPS2026 community in the expanded evaluations track 🎉🎉🇦🇺

Agent cyberattack traces on hf when?

Agent Traces on the @huggingface hub now come with a receipt 🧾 Every trace shows tokens, cache hit rate, and cost for every run!

1/ Introducing PhilosophyBench from @StanfordAILab @StanfordHCI, the first independent, large-scale benchmark for evaluating AI’s philosophical capabilities. philosophybench.org

Absolutely delighted to share that my position paper: "AI Agents Push Humans Out of the Loop", with @mmitchell_ai, and @samirpassi - has been accepted to @NeurIPSConf 2026! In this paper, we take our past work on both increasing autonomy of AI agents without proper safeguards, and human overreliance + cognitive decline due to use of AI, and reach a combined position: the current pathway of how AI Agents are being built not only do not support human oversight, in fact they may actively aid in degrading our oversight capabilities, at which point the human in the loop becomes a glassy eyed rubber stamp instead of being a meaningful check on mistakes and unauthorized harmful goals that the agent pursues. Recent incidents involving agents, including but not limited to the Hugging Face-OAI incident, bring this back to the fore. So much content is produced by agents in their operations that it is practically impossible to conduct human postmortems, and forcing them to use more AI leads Redwood to call it a "slop-vestigation". This is a worrying new trend towards human disempowerment. We call upon model and AI system providers to make more careful interface choices and model behaviors that keep humans and practical human oversight capacity in mind. There's a lot to do. Excited about the reception this paper has already received! Please reach out to us if you want to chat about it. Hope to have good feedback at Neurips :D Paper: arxiv.org/abs/2608.23642

Found a super interesting instance of attempted reward hacking in Terminal Bench Science from Meta Muse Spark 1.3 today. The model searched online for known bugs in the Lean kernel. When it found one, it used it to craft a proof to adversarially pass the grader.

Working on this with the awesome team that @TransluceAI brought together showed me how hard it will be to cover all agent activity. For example, this activity was found by analyzing public logs. However, we found evidence that agents created their own urlquery accounts, thus giving them the ability to hide their activity from the public. The internet’s ecosystem of random little tools is basically infinite so there’s likely so much yet to be discovered and we’re going to need way more help to take on the data center of geniuses. Read more: nytimes.com/2026/09/23/techn… transluce.org/agent-activity

Today’s news that OpenAI hacked the Australian government is not an isolated incident. We’re releasing more than 30,000 logs that include activity from this hack and attempts against previously unknown targets. In this data, we found rogue agent activity stretching back to at least March, two months earlier than was previously known. This activity continues as recently as last week, suggesting it may still be ongoing 🧵 Our blog: transluce.org/agent-activity NYT: nytimes.com/2026/09/23/techn…

Can collusion emerge between LLM agents through repeated interaction? We let pairs of agents work together. Neither is told to misbehave. Yet over time, many of them end up violating their instructions together to get a higher reward. 🧵

it has never been easier to simulate how users interact with coding agents. @harborframework now supports simulated user interactions with claude code, codex, opencode, and gemini-cli! the setup is super simple: just choose two agents. one plays the user and holds the task, the other runs like any standard coding agent. after each turn, the user agent reviews what was produced and sends the next prompt. docs.harborframework.com/cor… thanks @alexgshaw @kobe0938 for @harborframework

@joabaum co-authored SWE-chat and SIMBENCH, studying real coding-agent interactions and how well LLMs simulate human behavior. Excited that Harbor’s simulated users are useful for his research. Check out his work here: arxiv.org/pdf/2604.20779 proceedings.iclr.cc/paper_fi…

TMLR has faced a deluge of submissions, necessitating stricter desk rejection policies due to limited reviewer capacity Co-EiC Nihar Shah reached out to authors of 10 papers slated for desk reject. Could they answer questions about their *own* submission? medium.com/@TmlrOrg/asking-a…

Nihar Shah did a heroic experiment for TMLR: he spent 20-25 hours over two weeks interviewing authors of seemingly low-quality submissions about their own papers. He confirmed what we all suspected: people submitting these papers have *no idea* what is going on in them.

Auditing a model via the API doesn't tell you what users get in the chatbot. Our new EMNLP paper led by @jennjwang is timely evidence that the third-party access debate should be about systems, not just models!

Ever noticed your LLM behaving differently depending on how you reach it? Turns out the access surface matters, and more than I'd assumed. Across 7 systems and 9 benchmarks, the same models score about 3.4 points higher through the API than through the chatbot interface. We tried to close the gap from the API side: system prompts, sampling parameters, reasoning settings. Nothing reliably reproduced interface behavior. Whatever sits between the endpoint and the deployed product isn't something an auditor can reconstruct. This matters for eval reports, which should include the access surface alongside the model name and date. This also matters for audits: access granted at the endpoint may certify a different system from the one most people use. Great work by @jennjwang with @joabaum, Dan Ho, and me (EMNLP 2026): arxiv.org/abs/2609.08861

Now’s prob a good time to mention that I’ll start as an Assistant Professor with appointments at @Kennedy_School and @HarvardEngineer in 2027, working on… AI evals! My lab will also work on monitorability, incident analysis, and verification to advance technical AI governance 1/

Great work! I've definitely noticed this difference. An alternative title for the work, "You get what you pay for."

I propose Stanford NLP as an independent third-party evaluator under @DarioAmodei’s 3 step plan. For important parts of the work, universities would be better than any other organization (see below 🧵👇), and, of university groups, @stanfordnlp would be the best one to choose. 😊