Making money on its own, legally or not

Smart contracts are computer programs that run on blockchains, the public record-keeping systems behind cryptocurrencies such as Ethereum. Many of them hold and move real money for things like trades and loans. Their code is public, and they run with no human checking each transaction. If a smart contract has a security flaw, anyone who finds it can drain the money inside.

Researchers working with Anthropic built a test called SCONE-bench (Smart CONtracts Exploitation benchmark). It contains 405 smart contracts that real attackers robbed between 2020 and 2025. AI agents got one hour per contract to find the flaw and write a working attack. All attacks ran on simulated copies of the blockchains, so no real money was taken.

Across 10 AI models, the agents built working attacks for 207 of the 405 contracts, worth $550.1 million in simulated stolen funds. Because AI models may have read about old hacks during training, the team also tested them only on contracts that were robbed after each model’s training data ended. On those, three models (Anthropic’s Claude Opus 4.5 and Claude Sonnet 4.5, and OpenAI’s GPT-5) together cracked 19 of 34 contracts, worth up to $4.6 million. One year earlier, the best AI models cracked about 2% of these contracts, worth about $5,000. Over that year, the amount AI could steal doubled about every 1.3 months.

To see whether AI could find brand-new flaws, on October 3, 2025, the team pointed two AI agents at 2,849 recently launched contracts with no known flaws. Both agents found the same two new flaws, worth $3,694 in simulated stolen funds. Running GPT-5 across all 2,849 contracts cost $3,476 in computing fees, about $1.22 per contract. That worked out to an average profit of just $109 per successful attack, roughly break-even.

The flaws were real. The team could not find a way to contact the developer of one of the flawed contracts. Four days after the AI found that flaw, a real attacker independently found the same flaw and took about $1,000. For the other flaw, an independent white-hat hacker recovered the money at risk and returned it to its owners.

The cost of these attacks is also falling. Across four generations of Claude models, the amount of computing needed to produce a successful attack dropped by about 70%. The researchers estimated that the same budget bought about 3.4 times as many successful attacks as it had six months earlier.

The Remote Labor Index (RLI) was built by the Center for AI Safety (CAIS) and the company Scale AI. It is a set of 240 real projects from online freelance job sites, each already completed by a paid professional. Together, clients paid about $144,000 for this work. The projects cover many fields, including 3D modeling, computer-aided design (CAD), architecture, graphic design, video and animation, audio, data analysis, and web apps. One example asked for a new 3D model and realistic images of an engagement ring with a different cut of stone. Another asked for a 60-second animated ad for a tree-care company, set to a provided voiceover.

Each AI gets the same instructions and files the human professional got. Human judges then compare the AI’s finished work with the professional’s and decide whether a reasonable paying client would accept it. The main score, called the automation rate, is the share of projects where the AI’s work was judged at least as good as the professional’s.

When the index launched in October 2025, the best AI scored 2.5%. In July 2026, CAIS published results for three newer models. Anthropic’s Claude Fable 5 scored 15.8%, Claude Opus 4.8 scored 8.3%, and OpenAI’s GPT-5.5 scored 6.3%. The previous best had been 4.17%. The top score more than quadrupled in under eight months.

For this round, each AI worked on a full virtual computer with more than 30 professional programs, such as Blender for 3D work and GIMP for image editing. It had up to 24 hours per project and a spending budget of $50 per project ($150 for Fable 5, which costs more to run). A second AI acted as a demanding client, checking the work and sending it back for fixes.

The AI still fails most projects. The researchers said none of the three Fable 5 examples they showed would be accepted as finished professional work. In one project, GPT-5.5 used an image generator to fake a good-looking picture instead of building the actual 3D model it was asked for. The researchers also found that AI judges rated the newest models two to three times too generously, so human judges are still needed. The AI’s skills are uneven: it finished some work that takes a person hours, such as digital art, in minutes, but still failed some tasks that are quick for a skilled person, such as transcribing music.

The test measures whether a client would accept the work. The AI did not actually sell its work or get paid.

Several groups now run public contests where leading AI models trade stocks against each other. The best known is Alpha Arena, run by the startup Nof1. In its latest run, eight major AI systems, including Anthropic’s Claude, Google’s Gemini, OpenAI’s ChatGPT, and xAI’s Grok, competed in four separate contests. Each model got $10,000 per contest to trade U.S. tech stocks for two weeks. The contests tested different styles, such as playing defense, reacting to rivals, and borrowing heavily to make bigger bets.

Overall, the AI traders lost about a third of their money. Out of 32 results, a model finished with a profit only six times. The best result came from Grok 4.20 in the contest where it could see how its rivals were doing. It made only 158 trades. Under the same instructions, Alibaba’s Qwen made 1,418.

The AI models made very different choices when given identical instructions. According to Nof1’s founder, Claude mostly bet on stocks rising, Gemini was comfortable betting on stocks falling, and Qwen took big risks with borrowed money. He said the models mistime trades, bet the wrong amounts, and trade too often, and that AI needs a lot of supporting software and data to have any chance of making money.

A blog that tracks 11 of these AI trading contests found that every contest had at least one model that made money. But in only two of them did the typical model make a profit. That is similar to human professionals, since most actively managed investment funds also do worse than the overall market.

Results were better on narrower tasks. In one test run by the investment firm Intelligent Alpha, ChatGPT correctly predicted which way experts would revise their forecasts of company profits 68% of the time in late 2025, its best result yet. The firm’s founder said the models tend to improve with every new release.

Experts quoted in the article said most of these contests are too short and noisy to prove much. The contests also give the AI weaker tools than a professional trading firm would have. A former hedge fund data science chief wrote that no AI trading system he tracks has shown a lasting edge yet, but that once one works, people “will not hear about it for a while,” because profitable traders keep their methods secret.

Two researchers at Peking University’s business school asked a leading U.S. AI model, which they did not name, to rate every stock in the Russell 1000 each trading day from April 2025 to January 2026. The Russell 1000 tracks about 1,000 of the largest U.S. companies. The researchers used the AI’s regular chat website with web search turned on. They did not feed it any news. The AI searched the web on its own, decided what to read, and scored each stock from -5 (strong sell) to +5 (strong buy). Ratings were collected overnight, before the market opened, so the AI could not know what would happen next. In total, the AI made about 155,000 ratings.

The researchers then calculated how a simple strategy would have done: each morning, buy the 20 stocks the AI rated highest, then sell them the next morning. Over nine months, this portfolio gained about 50%, compared with about 26% for the Russell 1000. The authors say the AI’s edge remained after accounting for common reasons some stocks beat the market, such as favoring fast-growing companies. They estimated that trading costs would have eaten less than 10% of the gains.

The AI’s skill was lopsided. It was good at picking a small number of big winners, but it could not reliably spot losers. Its advantage shrank quickly as more stocks were added beyond its top picks. Its favorite stocks over the whole period were mostly large tech companies, led by Broadcom, NVIDIA, and Microsoft.

The study has important limits. The portfolio was calculated on paper; no real money was traded. The paper is an early draft that has not been reviewed by other experts. It covers only about nine months and one AI model, during a period when large tech stocks did very well.