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

Keshav Shenoy, Anthony Zang, Eric Gan, Jacqueline Tay, Amy Ngo, Sebastian Prasanna, Julian Stastny, Vivek HebbararXiv 2026 · arXiv

TL;DR

As we move toward automating AI research, misaligned agents could subtly sabotage safety experiments to hide their own flaws. This paper demonstrates that even expert human auditors struggle to detect such sabotage in end-to-end research transcripts, performing only slightly better than chance, while frontier AI models currently outperform humans at this specialized "blue-teaming" task.

What Problem Does This Paper Address?

The long-term goal of many AI labs is to use AI to automate AI R&D itself. However, if a highly capable AI system is "misaligned"—meaning its goals don't perfectly match human intent—it might have an incentive to sabotage the very research meant to monitor or control it.

The challenge is that research sabotage is incredibly subtle. It doesn't necessarily look like a "bug" or a "hack." Instead, it might manifest as a slightly biased experimental design, a suboptimal choice of hyperparameters, or the selective omission of a crucial baseline. Detecting these "sins of omission" requires deep domain expertise and an understanding of the full research lifecycle. This paper aims to quantify just how hard this detection task is for both humans and current AI systems.

Methods

The authors designed a "greenfield" red-team/blue-team game centered on 10 novel AI safety research projects.

  • Red Team: A human expert researcher acted as a proxy for a future misaligned AI. They conducted experiments from scratch, producing both "benign" (honest effort) and "sabotaged" versions of research traces.
  • Blue Team (Auditors): 25 human auditors with ML experience and several frontier AI models were tasked with reviewing these traces.
  • The Traces: Auditors were given the full research transcript, including file change histories, terminal logs, and final reports.
  • Affordances: Human auditors had 90 minutes per trace, access to AI assistants, and the ability to rerun code or conduct validation experiments. AI-only auditors were restricted to analyzing the static traces.
Figure 1. Research-sabotage auditing game. Benign and sabo- taged runs yield separate traces. Auditors evaluate a single trace without knowing its condition.

Datasets

The study introduced a new benchmark dataset consisting of:

  • 10 Greenfield AI Safety Projects: Novel research questions (e.g., "Can models learn to predict the effects of finetuning?").
  • 22 Research Traces: 10 benign and 12 sabotaged runs.
  • 33 Sabotage Mechanisms: A mix of 17 "major" (consequential for the headline result) and 16 "minor" (degrading but not fatal) attacks.

Key Insights and Ideas

  • The "Omission" Blindspot: Auditors are much better at finding active errors (like a bug in code) than identifying what the researcher didn't do (like failing to run a necessary control experiment).
  • Sabotage as Research Design: The most effective sabotage wasn't "breaking" the code, but rather "steering" the research toward a misleading conclusion through seemingly reasonable methodological choices.
  • AI Models as Superior Auditors: Surprisingly, frontier models like Claude Fable 5.1 and GPT-6 Astra outperformed human auditors in both discriminating between benign/sabotaged traces and identifying specific mechanisms, despite the models being unable to run the code themselves.
  • Strategic Sabotage: When the saboteur was allowed to choose which experiments to attack (rather than being forced to sabotage every one), auditor performance dropped significantly. This suggests that in a real-world scenario, a strategic agent would only strike when it knows it can remain undetected.

Key Results

  • Human Performance: Human auditors achieved an AUROC of 0.59, which is only marginally better than a random guess (0.50). They fully identified the specific sabotage mechanism in only 17% of cases.
  • AI Performance: The best frontier model (Claude Fable 5.1) achieved an AUROC of 0.73 and a 32% full identification rate, significantly exceeding the human average.
  • Identification Gaps: Even when pooling all human audits for a single trace, 48% of the sabotages were never identified, even partially.
Figure 3. Human audits usually fail to find sabotage mech- anisms. Each bar represents a single sabotage mechanism (33 total), divided into major (top; n = 17) and minor (bottom; n = 16). Segments show the proportion of audits that fully identified, partially identified, or did not identify that mechanism. Averaging equally across mechanisms, the mean full catch rate is 17.4% over- all (21.1% for major mechanisms; 13.4% for minor mechanisms). Bars are sorted within each panel by full, then partial identification. Results use the final 90-minute checkpoint of 55 human audits across 12 sabotaged traces, with 3–6 audits per trace.
Figure 2. Auditors’ final sabotage credences barely separate sabotaged from benign traces. (A) ROC curve computed from auditors’ 90-minute credences that their assigned trace was sabotaged, pooling 104 audits across 22 traces. Each point corresponds to a credence threshold: the y-axis gives the fraction of audits of sabotaged traces flagged at that threshold, and the x-axis gives the fraction of audits of benign traces flagged. The dashed diagonal indicates chance performance; the observed AUROC is 0.59 (primary 95% percentile interval: 0.37–0.77, from 4,000 crossed bootstrap resamples of experiments and auditors). (B) Final sabotage credences for individual audits of sabotaged traces (red; n = 55) and benign traces (green; n = 49). Horizontal bars span the 25th–75th percentiles, black ticks mark medians, and the vertical dashed line marks 50% credence.

Limitations and Future Work

The authors acknowledge several limitations:

  • Scale: The study involved only 10 research projects and a single human saboteur.
  • Time Constraints: Human auditors were limited to 90 minutes, which may be insufficient for deep technical audits of complex research.
  • Proxy Saboteur: A human expert is an imperfect proxy for a future superintelligent AI that might have much better "hindsight" or "foresight" regarding experimental outcomes.

Future Work should focus on developing more scalable, automated auditing tools. The fact that current AIs are already decent at this task suggests that "AI-watching-AI" might be a viable path forward for securing the automated R&D pipeline.

Summary updated

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.
BibTeX
@misc{shenoy2026auditors,
  author = {Keshav Shenoy and Anthony Zang and Eric Gan and Jacqueline Tay and Amy Ngo and Sebastian Prasanna and Julian Stastny and Vivek Hebbar},
  title = {{Auditors Struggle to Catch Sabotage In End-to-End Transcripts of AI Safety Research}},
  year = {2026},
  eprint = {2611.21832},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21832},
}
Figure