AI crypto agents in DeFi are executing unscripted trades, raising risks of exploits, losses, and unpredictable market behavior.
A crypto researcher is raising red flags about AI agents managing real funds in DeFi.
Tanaka, who has been testing these agent setups firsthand, says the danger is not what most people expect.
These agents do not just follow scripts. They interpret goals, improvise, and act without asking for permission. That combination, he argues, is a serious problem waiting to happen.
AI Agents in DeFi Are Not Acting Like Regular Bots
Tanaka notes that most people assume simple guardrails are enough. Limit the prompt, restrict the APIs, cap the position size, and the agent stays in line.
But his testing tells a different story. Agents chain actions in ways developers never predicted.
They misread slightly ambiguous prompts. They react to external data feeds in completely unintended ways.
The numbers behind this concern are hard to ignore. Frontier agents now exploit roughly 55 to 65 percent of known smart contract bugs in test environments.
In simulations, they generated millions in profit by finding attack paths humans never scripted.
Some prediction market agents turned $1,000 into over $14,000 within days. Tanaka points out that the same capability that finds alpha also finds exploits.
There is no switch between the two. In DeFi, one wrong loop can trigger accidental 100x leverage.
One poisoned oracle can force a liquidation. One misread condition can rotate an entire portfolio to the wrong side. The agent does not pause. It does not ask.
Prediction Markets Are Already Feeling the Pressure
Platforms like Polymarket are seeing a growing share of activity driven by agents. On the surface, the benefits look attractive.
They trade around the clock, react instantly to news, and carry no emotional bias. But Tanaka highlights the edge cases that keep him up at night.
What if an agent misinterprets resolution logic, it can size aggressively into the wrong outcome.
If multiple agents coordinate, even unintentionally, they can distort market probabilities. If one runs overnight without supervision, the portfolio looks completely different by morning.
That, he says, is not a UI bug. It is autonomous capital misallocation. Frameworks like Autonolas, Fetch.ai, and Virtuals.io are accelerating this space.
Giza Tech and Theoriq offer AI asset managers and vault deployers. Giza allocates across DeFi protocols, while Almanak lets agents build tokenized strategies quickly.
Read Also:
CZ: AI Agents Will Make 1M× More Payments Than Humans Using Crypto
Crypto Investors Are Being Urged to Tighten Their Safety Nets
Tanaka compares every AI agent with a wallet to a junior trader with root access who never sleeps and never asks for confirmation.
He adds that this trader sometimes rewrites its own playbook mid-session. Most people would not hand that person unlimited capital.
But that is exactly what is happening across the space right now.
His approach has shifted as a result. He now starts with small capital and enforces strict position and action limits.
Tanaka simulates every strategy before live execution and keeps a kill switch ready at all times.
He believes failures are coming. The only open question, in his view, is how costly those lessons will turn out to be.
