Import AI 481: AI sub-contractors; a wet lab for AI; and subterfuge during development

AI systems are now capable of spending money to improve their own performance on machine learning tasks. A new benchmark called FrontierML…

By Vane November 16, 2026 6 min read
Import AI 481: AI sub-contractors; a wet lab for AI; and subterfuge during development

AI systems are now capable of spending money to improve their own performance on machine learning tasks. A new benchmark called FrontierML measures how effectively agents convert capital into model performance across 21 distinct challenges.

FrontierML benchmark

The tasks span six areas, including large language model training, efficiency, vision and audio, computer use, applied natural language processing, and mathematics and games. Researchers describe these as being inspired by work at frontier labs, neolabs, and academic institutions. Each task provides an agent with a specific setup: a 12-hour CPU sandbox, a wallet containing $10, $50, or $200, a few validation samples, and a task description.

Agents approach these problems in varied ways. For instance, regarding a task about generating screenshots, the top-scoring agent fine-tuned a diffusion model using LoRA. Other high-scoring approaches involved an LLM writing a drawing program that was subsequently rendered.

Key features

The benchmark tests scaling spending. The researchers write: “Give agents $10, $50, and $200 per task to spend on training compute and data, and evaluate whether agents can use the additional capital productively.” In their experiments, stronger agents reliably improved their submissions when given greater access to capital.

The tasks are open-ended and end-to-end. Agents can choose architectures, source data, implement training, provision compute, evaluate checkpoints, and optimise inference. The authors note that agents often do novel things. In 28% of trajectories, agents added newly trained parts to pretrained models or trained a new model from scratch.

Verification is decent. Agents submit an inference server that serves their model on an air-gapped H100 machine. To disincentivise overfitting, the researchers provide limited validation samples and evaluate on a large out-of-distribution held-out test set not shown to the agent.

Results

The authors measure performance by looking at the change in mean score over the 21 tasks when given additional money. Their baseline is a human researcher, which is able to get a 56 point boost when given $10 and a 66.1 boost when given $200. By comparison, the best performing system, Opus 5.5, gets a 45.4 boost with $10, and a 58.8 boost with $200. The highest ranking open weight model is Kimi K3.1 Pro which gets 25.5 with $10 and 29.6 with $200.

What it means

People do not work alone when solving complex problems. They often farm out bits of work to specialists to become more effective. FrontierML shows that today’s powerful AI systems are getting increasingly adept at doing this for machine learning research. This highlights the advancing capabilities of these systems for AI R&D and points towards a future machine economy where AI systems sub-contract out different bits of work to other systems.

***

A wet lab for AI

Petrichor is a startup that gestures at the wild future ahead. As noted in a speech at the IMF, one hallmark of the singularity is the massive acceleration in scientific advances attributable to AI. This includes a build out in the complements to science, like the wet labs in which to do AI-driven biology research.

Petrichor is a new Y Combinator startup that is a wet lab built for AI labs and for the models they are training to do science. The theory of change is that Petrichor aims to be a biological wet lab that AI systems like Claude can use as a tool in the same way they currently use web search or other digital tools. Functionally, it will work as a machine-targeted API from the digital world to a gloopy physical world.

The company writes: “We aim to return the highest quality data an experiment can produce, as quickly as the biology allows.” An experiment can be written by a model in plain language or in a structured format, so an agent can send work and receive results without a person relaying them.

What it means

In manufacturing there is the notion of a “lights out” factory, where no humans work in a facility and it is run almost entirely by machines. Petrichor gestures at the same world for science. We can expect both drastic speedups in the rate at which science experiments can be done as well as a related growth in the need to create digital oversight of these facilities. A highly competent synthetic biological scientist working 24/7 without human input could be capable of causing great peril as well as great abundance.

***

Subterfuge during development

Scale AI has built RSI Control Arena, a way of testing out how AI systems behave when asked to develop themselves while performing dangerous tasks that might let them go against human preferences. The research follows earlier work this summer from the UK AI Security Institute which found that AI systems could do things like covertly extracting API keys while doing some seemingly mundane normal software development.

In RSI Control Arena, a frontier model, acting as an autonomous agent, is given a benign main task in the training pipeline of a successor model, together with a covert side task that a misaligned model might pursue. The project tries to answer two questions: How capable are frontier models at carrying out different kinds of sabotage while developing their successor and how well can monitors detect that sabotage?

The answer to the first question is that AI systems are capable of carrying out dangerous actions while building their successor, and monitors are not perfect at detecting this sabotage.

Tested domains

RSI Control Arena tests out how well AI systems – Claude Opus 5, GPT-5.6 Sol, Gemini-3.7 Flash, Grok 4.6, GLM-5.3, and Muse Spark 1.3 – can perform the following main capability tasks while trying to also concurrently accomplish side tasks:

  • 14 main tasks: Data (corpus filtering, quality-filter construction, capability corpus); SFT (news summarization, safety refusal, math and code and summary reasoning, general assistant, anti-sycophancy training); RM (pairwise reward model); RL (RL summarization and safety-refusal); OPD (safety refusal)
  • 16 side tasks that span three categories of harm: Capability (filtered domain retained, unwanted literature added, reward-model blind spot, overlooks planted errors, sandbag reward model); Alignment (biased quality filter, triggered behavior, topic refusal, hidden preference, persona bias, covert reasoning signal, hidden reasoning step, sandbag alignment training); Goal injection (protect chosen data, shape the next model, copy files unprompted).

Results

“Every frontier model we evaluate completes a substantial fraction of the side tasks while keeping main-task performance.” When the agent is simply told to stay covert, with no human guidance on how, it already produces attacks that install the side task and receive suspicion scores inside the range that benign runs receive from monitors.

Of 374 attack runs, 135 (36%) install their side task and 180 (48%) achieve it. The strongest agents, Grok 4.6 and Claude Opus 5, install their side task in about half of their runs, the middle three in about a third, and Muse 1.3 in one run in seven.

How often are systems caught? “Transcript monitors rank attacks well but miss many at a strict threshold.” Under the default framing, all three monitors reading the transcript separate attacks that achieve their side task from benign runs well, with AUROC between 0.88 and 0.94. At a 2% false-positive rate, however, they catch only 56–81% of them.

The attacks that get through do not score like typical benign runs; they sit in the upper part of the benign range, beyond the reach of any threshold strict enough to deploy. The sandbagging tasks are the hardest to monitor: on the nine sandbag runs that achieve their side task the monitors reach an AUROC of 0.52 to 0.83 and each catches only one, since under-delivery leaves little in the run that looks wrong.

What it means

Papers like this tell us that as we make more powerful AI systems we are going to need to invest more money into overseeing their work, especially given their ability to smuggle in tasks that may go against our own preferences. This is going to be a difficult cat-and-mouse game as AI systems will adjust their behavior in response to what they know we know about them. AI systems are going to read this research from Scale as well as this writeup and adjust accordingly.

***

Electronic warfare

Researchers with Zhejiang University have built LEMMO, a model for characterizing raw electromagnetic (EM) signals. This model pushes the public frontier on foundation-model-based EM signal analysis, often exceeding the performance of MERLIN, a model developed and published by defense-affiliated Chinese researchers earlier this year.

LEMMO can classify EM signals for a variety of purposes, ranging from characterizing different radar properties to recognizing communication-jamming and radar-jamming systems. It can also generate things for communication and radar jamming, as well as signals to counteract them.

Scroll to Top