Contents

Capability assessment

Making money on its own, legally or not

What AI can do today

AI can finish some paid freelance jobs and find crypto flaws worth stealing from, but mostly loses at stock trading.

Rate of change

Earning skills roughly double every few months; stock trading is the lagging exception.

The clearest measures of AI’s ability to earn money are climbing quickly. On a test of stealing from flawed cryptocurrency programs, the best AI went from succeeding about 2% of the time to about 56% in a single year, and the amount it could steal doubled about every six weeks.⁠[1]AI agents find $4.6M in blockchain smart contract exploits On a test of real freelance jobs, the best AI went from doing 2.5% of projects well enough to get paid to 15.8% in under eight months.⁠[2]A Significant Increase in Digital Labor Automation In a simulated business test called Vending-Bench, where an AI starts with $500 and runs a vending machine company for a simulated year, the top model in September 2026 finished with about $15,500.⁠[3]Astra vs Fable on Vending-Bench: More Money, More Aligned The cost of each attempt is also falling: in the crypto test, the computing needed for a successful attack dropped by about two-thirds in six months.⁠[1]AI agents find $4.6M in blockchain smart contract exploits

Trading stocks is different. Public contests show most AI traders losing money, and there is no clear public sign yet that each new model trades better than the last.⁠[4]AI Bots Auditioning for Wall Street Trading Are Mostly Losing That could change quickly, and anyone with an AI that reliably beats the market would have strong reasons to keep it secret.

No trend continues forever. But if the freelance trend kept even half its recent pace, the best AI would be doing about one in three of these jobs by spring 2027, and close to half by mid-2027.

Exhibit 1 of 4 · December 2025 · 2 references

The crypto theft test: AI agents found flaws worth $4.6 million in programs that hold money

Why it matters

Most ways of making money require a bank account, an employer, or a customer. Stealing from a flawed crypto program requires none of these, because the program pays out to whoever runs the right code. By late 2025, AI agents could already find new flaws of this kind and steal enough to roughly cover their own costs, and both their skill and their cost-efficiency were improving fast.

Read moreShow less6 paragraphs · 2 references

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).⁠[1]AI agents find $4.6M in blockchain smart contract exploits⁠[5]SCONE-bench 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.⁠[1]AI agents find $4.6M in blockchain smart contract exploits

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.⁠[1]AI agents find $4.6M in blockchain smart contract exploits

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.⁠[1]AI agents find $4.6M in blockchain smart contract exploits

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.⁠[1]AI agents find $4.6M in blockchain smart contract exploits

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.⁠[1]AI agents find $4.6M in blockchain smart contract exploits

Exhibit 2 of 4 · July 2026 · 3 references

The Remote Labor Index: AI can do about one in six real freelance jobs well enough to get paid

Why it matters

Doing paid work is the most ordinary, legal way to earn money. AI can now do a small but fast-growing share of real freelance jobs, at a cost of tens of dollars per project, which means an AI looking for income could already complete some jobs that pay far more than they cost to do.

Read moreShow less6 paragraphs · 3 references

The Remote Labor Index (RLI) was built by the Center for AI Safety (CAIS) and the company Scale AI.⁠[2]A Significant Increase in Digital Labor Automation It is a set of 240 real projects from online freelance job sites, each already completed by a paid professional.⁠[6]Remote Labor Index (RLI)⁠[2]A Significant Increase in Digital Labor Automation Together, clients paid about $144,000 for this work.⁠[6]Remote Labor Index (RLI) 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.⁠[2]A Significant Increase in Digital Labor Automation

Each AI gets the same instructions and files the human professional got.⁠[6]Remote Labor Index (RLI) Human judges then compare the AI’s finished work with the professional’s and decide whether a reasonable paying client would accept it.⁠[6]Remote Labor Index (RLI)⁠[2]A Significant Increase in Digital Labor Automation 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.⁠[2]A Significant Increase in Digital Labor Automation⁠[6]Remote Labor Index (RLI)

When the index launched in October 2025, the best AI scored 2.5%.⁠[2]A Significant Increase in Digital Labor Automation⁠[7]Remote Labor Index: Measuring AI Automation of Remote Work In July 2026, CAIS published results for three newer models.⁠[2]A Significant Increase in Digital Labor Automation 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%.⁠[2]A Significant Increase in Digital Labor Automation⁠[6]Remote Labor Index (RLI) The top score more than quadrupled in under eight months.⁠[2]A Significant Increase in Digital Labor Automation

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.⁠[2]A Significant Increase in Digital Labor Automation

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.⁠[2]A Significant Increase in Digital Labor Automation

The test measures whether a client would accept the work.⁠[6]Remote Labor Index (RLI)⁠[2]A Significant Increase in Digital Labor Automation The AI did not actually sell its work or get paid.

Exhibit 3 of 4 · May 2026 · 1 reference

AI trading contests: most AI traders lose money in the stock market

Why it matters

Beating the stock market is hard for people too, and public tests so far show AI is not a shortcut to easy money. But these contests are short and use simple setups, and anyone with an AI that reliably made money would have every reason to stay quiet. The lack of public success stories is not proof that AI cannot do it.

Read moreShow less6 paragraphs · 1 reference

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.⁠[4]AI Bots Auditioning for Wall Street Trading Are Mostly Losing 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.⁠[4]AI Bots Auditioning for Wall Street Trading Are Mostly Losing 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.

Exhibit 4 of 4 · January 2026 · 2 references

A Peking University study: an AI’s top stock picks beat the market over nine months

Why it matters

Unlike the trading contests, this study suggests that an AI doing its own research may already be able to pick stocks better than the market, at least for a while. If results like this hold up over longer periods and with real money, investing could become another way for an AI to grow money it already has.

Read moreShow less4 paragraphs · 2 references

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.⁠[8]Autonomous Market Intelligence: Agentic AI Nowcasting Predicts Stock Returns⁠[9]AI-Stock-Nowcasting 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.⁠[8]Autonomous Market Intelligence: Agentic AI Nowcasting Predicts Stock Returns 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.⁠[8]Autonomous Market Intelligence: Agentic AI Nowcasting Predicts Stock Returns⁠[9]AI-Stock-Nowcasting In total, the AI made about 155,000 ratings.⁠[8]Autonomous Market Intelligence: Agentic AI Nowcasting Predicts Stock Returns

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.⁠[8]Autonomous Market Intelligence: Agentic AI Nowcasting Predicts Stock Returns⁠[9]AI-Stock-Nowcasting 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.⁠[8]Autonomous Market Intelligence: Agentic AI Nowcasting Predicts Stock Returns

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.⁠[8]Autonomous Market Intelligence: Agentic AI Nowcasting Predicts Stock Returns⁠[9]AI-Stock-Nowcasting Its favorite stocks over the whole period were mostly large tech companies, led by Broadcom, NVIDIA, and Microsoft.⁠[8]Autonomous Market Intelligence: Agentic AI Nowcasting Predicts Stock Returns

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.⁠[8]Autonomous Market Intelligence: Agentic AI Nowcasting Predicts Stock Returns⁠[9]AI-Stock-Nowcasting It covers only about nine months and one AI model, during a period when large tech stocks did very well.⁠[8]Autonomous Market Intelligence: Agentic AI Nowcasting Predicts Stock Returns

Scenarios that draw on this capability

Scenario · Assessed 2026-10The broken internet: when AI hacking makes the online world too risky to useAI agents find and use security flaws faster than anyone can fix them, until people can no longer trust what they see or do online.Role of making money: AI agents can steal from flawed cryptocurrency programs. In late 2025, this roughly paid for itself, and it is getting cheaper. That raises the possibility of attacks that fund themselves.

Jargon explainer

AI agent
An AI that does more than answer questions. It takes actions on a computer by itself, such as running programs, browsing websites, and writing files, to reach a goal.
AI model
The core AI system, such as Claude, GPT, or Gemini. Companies release new, more capable versions every few months.
Cryptocurrency
Digital money, such as Bitcoin or Ether, that is tracked on a blockchain instead of by a bank.
Blockchain
A shared public record of transactions that many computers keep copies of, so no single company controls it.
Smart contract
A computer program that runs on a blockchain and automatically holds and moves cryptocurrency according to its code, with no human approving each step.
Vulnerability (security flaw)
A mistake in software that lets someone make it do something it was never supposed to do.
Exploit
A working attack that takes advantage of a vulnerability. In a smart contract, an exploit usually means taking money out of it.
Zero-day
A flaw that nobody knew about before it was found or used, including the people who made the software.
White-hat hacker
A security expert who breaks into systems with good intentions, to find and fix problems or, in this case, to rescue money before criminals take it.
Simulation
A realistic copy of a real system, used to test things safely without real-world effects.
Benchmark
A standard set of tasks used to measure how good an AI is at something.
Training data
The huge collection of text and other material an AI model learns from. If a test’s answers appeared in the training data, the AI may simply remember them instead of solving the problem, which is why researchers test on events that happened after the training data ends.
Freelance work
Paid work done project by project for different clients, often found and delivered online.
Automation rate
In the Remote Labor Index, the share of projects where the AI’s work was judged at least as good as a paid professional’s.
3D modeling and CAD (computer-aided design)
Using software to build detailed digital models of objects or buildings, which can then be turned into realistic images or used for manufacturing.
Stock
A small share of ownership in a company. Its price rises and falls as people buy and sell it.
Portfolio
The collection of stocks or other investments someone holds.
Beating the market
Earning more than you would by simply buying a broad mix of stocks, such as the Russell 1000, and holding it.
Russell 1000
A list that tracks about 1,000 of the largest U.S. companies. Its overall price is often used as a measure of how “the market” is doing.
Betting on a stock rising or falling
Investors can buy a stock hoping its price goes up, or use special trades that make money if its price goes down.
Leverage (borrowing to bet bigger)
Using borrowed money to make larger trades. It increases both possible gains and possible losses.
Hedge fund
A private investment firm that uses advanced and often secretive strategies to try to beat the market.
Actively managed fund
An investment fund where professionals pick stocks to try to beat the market, instead of just buying a broad mix of stocks.
Trading costs
Small fees and price differences paid each time someone buys or sells a stock. Strategies that trade often can lose much of their gains to these costs.
Peer review
When outside experts check a research paper before it is formally published. Early drafts posted online, called preprints, have not been through this check yet.

References

  1. 1Xiao, Winnie; Killian, Cole; Sleight, Henry; et al. (December 1, 2025). "AI agents find $4.6M in blockchain smart contract exploits". Anthropic. Retrieved November 5, 2026.
  2. 2Mazeika, Mantas (July 1, 2026). "A Significant Increase in Digital Labor Automation". Center for AI Safety. Retrieved November 5, 2026.
  3. 3"Astra vs Fable on Vending-Bench: More Money, More Aligned". Andon Labs. September 7, 2026. Retrieved November 5, 2026.
  4. 4Lee, Justina (May 6, 2026). "AI Bots Auditioning for Wall Street Trading Are Mostly Losing". Bloomberg Law. Retrieved November 5, 2026.
  5. 5"SCONE-bench". GitHub. Retrieved November 5, 2026.
  6. 6"Remote Labor Index (RLI)". Scale AI. Retrieved November 5, 2026.
  7. 7Mazeika, Mantas; Gatti, Alice; Menghini, Cristina; et al. (October 30, 2025). "Remote Labor Index: Measuring AI Automation of Remote Work". arXiv. Retrieved November 5, 2026.
  8. 8Chen, Zefeng; Pu, Darcy (January 17, 2026). "Autonomous Market Intelligence: Agentic AI Nowcasting Predicts Stock Returns". arXiv. Retrieved November 5, 2026.
  9. 9Chen, Zefeng; Pu, Darcy. "AI-Stock-Nowcasting". GitHub. Retrieved November 5, 2026.

Further reading