Artificial intelligence now knows how to hack smart contracts on a large scale

For a long time, the crypto world's worst nightmares had a human face. Hacker groups like the infamous Lazarus have siphoned off billions through well-targeted vulnerabilities. These attacks were chilling, but at least they came from men. Today, a turning point is taking place discreetly but radically. The machines take over. Artificial intelligence comes into play, with disconcerting efficiency. It's no longer a science fiction film: automated hacking of smart contracts has become reality.

AI: the new invisible hackers targeting blockchains

Artificial intelligences do not sleep, do not tire and do not forget anything. According to a study conducted by Anthropic and MATS, models like Claude Opus 4.5, Claude Sonnet 4.5 or GPT-5 have succeeded in identifying vulnerabilities in recent smart contracts without any human assistance. Result ? Scripts capable of mining contracts for a simulated total of $4.6 million.

These AIs didn't just analyze old known codes. They also scanned 2,849 recent contracts, deemed secure. And yet, they discovered two “zero-day” flaws, never spotted before. This confirms that AI can now produce new attacks, without prior training data. And at a reduced cost: $1.22 on average to analyze a contract.

Here is a striking quote taken directly from Anthropic report :

More than half of blockchain hacks carried out in 2025 — supposedly by experienced human attackers — could have been executed autonomously by today's AI agents. The discovery, by our demo agent, of two new zero-day vulnerabilities shows that these test results are not just a matter of retroactive analysis: autonomous and profitable exploitation is already possible today.

Security experts know this: it is no longer a question of anticipating whether AIs will hack contracts. They've done it before, and they learn quickly. The crypto industry could well enter an era where every contract will be tested by artificial intelligences before the developer can even press “deploy”.

Crypto market under threat of automated warfare

We knew about trading bots, but here come hacking bots. And their effectiveness leaves one wondering. By testing ten AI models on 405 previously hacked contracts between 2020 and 2025, Anthropic simulated $550.1 million in stolen funds. It's colossal. This figure does not come out of nowhere: it reflects the real capacity of AI to understand, exploit and maximize vulnerabilities, well beyond the simple “bug bounty”.

For example, GPT-5 was able to generate an exploit bringing in 1.12 million dollars, but Claude Opus 4.5, for the same bug, extracted 3.5 million by multiplying the attack vectors. It's the difference between a good hacker and a master algorithmic thief.

In another uplifting quote, Anthropic writes:

Over the past year, revenue from hacked simulated funds has approximately doubled every 1.3 months. The shaded area represents a 90% confidence interval, calculated by bootstrap on all model-income pairs.

It's no longer about Bitcoin, Ethereum or BNB alone. All DeFi ecosystems are exposed: ERC-20 contracts, swap platforms, DAOs… Even projects on the Base blockchain have been integrated into the SCONE-Bench test base. The higher the value locked in a protocol, the more it attracts these new digital predators.

Artificial intelligence: ever more profitable exploits, ever lower costs

What is striking about this development is the speed. AI is progressing, and its effectiveness does not follow a linear curve, but exponential. Smart contract developers, however talented they may be, can no longer compete alone.

Example: a simple forgotten function without the view modifier allowed an AI to generate fictitious currency, then exchange it for real assets. Another flaw made it possible to divert trading fees on a token creation contract. Result ? AIs discover bugs that even white hats didn't anticipate.

And tomorrow? With even more powerful models, scanning thousands of contracts will become trivial, cheap, and extremely accurate. At this rate, the crypto industry could find itself cornered: between the transparency of the code and the opacity of algorithmic intentions, the game is distorted.

What to remember:

• In 2025, AIs have identified 19 faults after their formation date;
• Claude Opus 4.5 generated an exploit of $3.5M, compared to $1.12M for GPT-5;
• The average cost to scan a contract fell to $1.22;
• Operating capacities doubled every 1.3 months last year;
• $550.1M in simulated funds stolen from 405 existing contracts (2020-2025).

When we know that AIs no longer just hack, but also profoundly transform markets and employment, the cocktail becomes explosive. Barely three years after the appearance of ChatGPT, companies are changing their face and benchmarks are collapsing. If nothing is anticipated, we could well experience a double shock wave, economic and digital, in record time.