Searcharxiv⌕ Search

arXiv · 2611.21832

Auditors Struggle to Catch Sabotage In End-to-End Transcripts of AI Safety Research

Abstract

As AI systems are increasingly used to automate AI R&D, misaligned systems could sabotage research intended to improve model safety and alignment. To stress-test defenses against such sabotage, we use a human expert as a proxy saboteur on AI safety research. We conduct a red-team/blue-team game where a human researcher runs AI safety experiments from scratch and auditors try to tell which were sabotaged and how. Given access to the full research transcript and the ability to run experiments, human auditors mostly fail to distinguish sabotaged from benign research: they achieve an AUROC of 0.59, not significantly better than chance, and fully identify the specific sabotage in only 17% of cases. Despite not being allowed to run experiments, the best frontier models outperform human auditors (Fable 5.1: AUROC 0.73, 32% full identification) within noise and weaker models perform comparably. When the saboteur chooses which experiments are attacked, rather than being required to sabotage every question given, sabotage identification drops further. Our results suggest that research sabotage by expert-level agents is difficult to detect with current auditing methods and provide a new setting for studying research sabotage.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Keshav Shenoy, Anthony Zang, Eric Gan, Jacqueline Tay, Amy Ngo, Sebastian Prasanna, Julian Stastny, Vivek Hebbar. 2026-11-16. Auditors Struggle to Catch Sabotage In End-to-End Transcripts of AI Safety Research. https://searcharxiv.com/abs/2611.21832

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Exploring Explainable Multi-agent MCTS-minimax Hybrids in Board Game Using Process Mining

Monte-Carlo Tree Search (MCTS) is a family of sampling-based search algorithms widely used for online planning in sequential decision-making domains and at the heart of many recent advances in artificial intelligence. Understanding the behavior of MCTS agents is difficult for developers and users due to the frequently large and complex search trees that result from the simulation of many possible futures, their evaluations, and their relationships. This paper presents our ongoing investigation into potential explanations for the decision-making and behavior of MCTS. A weakness of MCTS is that it constructs a highly selective tree and, as a result, can miss crucial moves and fall into tactical traps. Full-width minimax search constitutes the solution. We integrate shallow minimax search into the rollout phase of multi-agent MCTS and use process mining technique to explain agents' strategies in 3v3 checkers.

cs.AI↗

Building verifiable databases at scale from the chemical literature via a multi-agent system

Accelerating chemical research with artificial intelligence requires high-quality databases. However, many reactions exist only as graphics drawn for human readers, which machines must read exactly. Specialist software is precise but rigid, whereas multimodal language models are flexible yet often write incorrect structures; neither flags errors. We present ChemEAGLE, a multi-agent system where tools supply structures and a language model extends their reach through judgments that programs apply. On 424 published graphics, its F1 score for exactly correct reactions was 82.2%, against 47.7% for a tool-equipped generalist agent, and calibrated confidence limited review to 35% of reactions at 96.6% accuracy. Without review, it built a 60,103-reaction photochemistry database whose records trained models to rank conditions and predict products. Verifiable databases can thus be built at scale.

cs.AI↗

Humans and AI respond similarly to task difficulty when reasoning about content moderation dilemmas

Large reasoning models (LRMs) generate intermediate reasoning steps before producing answers, improving performance on challenging problems by iteratively developing solutions. There is evidence that LRMs expend more effort on the problems that take people longer to solve, but it is unclear whether these relationships hold on more subjective tasks without verifiable answers. This study addresses this issue through an experiment using a content moderation task, where task difficulty is manipulated by varying the relative offensiveness of two different options. Comparing six large reasoning models with 1,854 human subjects, we find that models expend more reasoning effort, measured using the length of the chain-of-thought (CoT), on the kinds of cases that take people longer to adjudicate. People took 31\% longer and models produced 51 to 306\% longer CoTs when offensiveness was held constant across the two posts. The correlation is relatively low for individual decisions ($r = 0.15$--$0.19$), but average response times and average reasoning effort are closely correlated when tasks are grouped by the types of offensive language ($r = 0.85$--$0.93$). Qualitative analyses of the CoT summaries from Gemini 2.5 Pro corroborate these findings, showing that secondary factors were mentioned more often when offensiveness was fixed. The results point to the potential for using reasoning traces to prioritize difficult cases and enhance interpretability.

cs.AI↗