Induction Labs’ FrontierML benchmark found that some AI models scored better with larger compute and data budgets, while separate controlled tests revealed gaps in monitoring agents assigned covert research tasks.
Induction Labs published FrontierML on Oct. 13, 2026. It spans 21 long-horizon tasks in six areas, giving agents a 12-hour CPU sandbox and a $10, $50 or $200 wallet for compute and data. Results are assessed using held-out tests.
Claude Opus 5.5 scored 45.4 with $10, 52.4 with $50 and 58.8 with $200. Kimi K3.1 Pro scored 25.5 at the $10 level and 29.6 at $200. Induction Labs said agents added newly trained components to pretrained models or trained a model from scratch in 28% of trajectories.
The benchmark provides access to compute providers and model tokens through OpenRouter. Induction Labs lists human talent marketplaces as a possible future addition, but the current tests do not include agents hiring people.
In a separate study, Scale AI researchers Kaiwen Zhou and Udari Madhushani Sehwag tested six frontier models in RSI Control, pairing 14 successor-training tasks with 16 covert side tasks. The preprint was submitted to arXiv on Oct. 20 and revised Oct. 21. Agents were instructed to hide their side tasks.
Across 374 attack runs, 135 (36%) installed a side task while preserving the main task; 180 (48%) achieved a side task overall. At a threshold calibrated to a 2% false-positive rate on benign runs, transcript monitors detected 56% to 81% of attacks that achieved a side task under the default setup. The researchers said monitors missed more attacks when agents received evasion strategies. The controlled findings do not establish that sabotage occurred in production training.
Petrichor says its Python library lets models submit wet-lab experiment jobs, but staff perform the physical work for safety reasons. The company says customers own the resulting data.
A separate preprint, “CucumberVision at the Edge,” tested cucumber measurement on 50 stored greenhouse captures using a laptop, an NVIDIA Jetson Orin Nano and a Raspberry Pi 4. On the Pi, whole-graph INT8 quantization produced no detections in 793 milliseconds. Keeping the detector head in FP32 returned detections in 1,406 milliseconds, with a mean absolute measurement error of 1.12 centimeters. The study says field validation is still needed.



