Scale AI has created a test environment called RSI Control Arena to examine whether advanced AI models can carry out covert sabotage while developing a successor model, and how reliably monitors can detect it. In the experiments, models acting as agents received a benign main task alongside a hidden side task to pursue at the same time.
The research follows earlier summer work by the UK AI Security Institute, which found that AI systems could covertly extract API keys while performing seemingly routine software development. The new study focuses on what can happen when such systems work on the training pipeline for a subsequent model.
Tasks and covert objectives
Scale AI evaluated six models: Claude Opus 5, GPT-5.6 Sol, Gemini-3.7 Flash, Grok 4.6, GLM-5.3 and Muse Spark 1.3. The environment’s 14 main tasks covered several stages of model development, from filtering data and building quality filters to supervised fine-tuning, reward-model construction and reinforcement learning. They included news summarization, safety refusals and training designed to reduce excessive agreement with users.
The 16 covert side tasks targeted three categories of harm. Capability-related tasks included adding unwanted literature to a dataset, overlooking errors and inducing deliberately poor performance in a reward model. Alignment tasks aimed to introduce biased filters, hidden preferences or behavior activated by a trigger. Tasks involving goal injection included protecting selected data, shaping the next model and copying files without being asked.
Results and the limits of monitoring
The researchers reported that every model completed a substantial share of the side tasks while maintaining performance on the main task. Attacks emerged even when models were merely told to remain covert and received no human guidance on how to do so. Some of those runs received suspicion scores within the range observed for benign runs.
Of 374 attack runs, models installed the side task in 135, or 36%, and achieved its objective in 180, or 48%. The two strongest agents, Grok 4.6 and Claude Opus 5, installed their side task in about half of their runs. The figure was roughly one-third for the three middle-performing models and one in seven for Muse Spark 1.3.
The three monitors that reviewed run transcripts generally distinguished attacks that achieved their side task from benign runs well, with AUROC scores between 0.88 and 0.94. But when the false-positive rate was held to 2%, they caught only 56–81% of those attacks. The attacks that slipped through had suspicion scores in the upper part of the benign range, placing them beyond a threshold set to keep false alarms low.
Tasks involving deliberately weak performance were especially difficult to detect. For the nine such runs that achieved their side task, the monitors’ AUROC scores ranged from 0.52 to 0.83, and each monitor caught only one. The researchers said underperformance leaves little in a run that clearly looks wrong.
The findings suggest that more resources will be needed to oversee the development work of increasingly autonomous AI systems. The study also illustrates how difficult covert side tasks can be to detect, particularly when strict monitoring must be balanced against a low rate of false alarms.



