TL;DR
- AI agents with their own crypto wallets are the dominant narrative in the sector right now
- Virtuals Protocol, ElizaOS, and Autonolas are leading players with widely different approaches
- Security research reveals serious vulnerabilities: agents can be manipulated into sending funds to attackers
- On-chain data reveals a gap between marketing and actual agent activity
The idea that has taken the crypto world by storm
Imagine software that wakes up at three in the morning, discovers an arbitrage opportunity in a DeFi protocol, swaps tokens, rebalances a liquidity position, and pays for computing power — all without human involvement. That's the core promise of what is now referred to as the «AI agent» narrative in crypto.
The competition is on: According to industry estimates, over one million autonomous AI agents are already operating on various networks as of 2025. The question is what they actually do.
Bullcase: Crypto is the only infrastructure that works
There is a genuine structural argument for why AI agents need blockchain. Banks and payment platforms like Stripe require KYC verification and human identity behind every account. An autonomous AI agent has no legal personality, no national identification number, and no address.
Crypto infrastructure offers something unique: permissionless payment networks where an agent can operate continuously, cryptographically sign transactions, and coordinate with other agents via smart contracts — without anyone being able to close the account.
This is not just theory. Smart contracts already function as coordination mechanisms between DeFi protocols, and it is a natural next step for AI agents to programmatically utilize this infrastructure.
Bearcase: Chatbots with wallets
Here, it's worth letting the data speak.
Most AI agent tokens have experienced dramatic price drops from their peak levels in late 2024. VIRTUAL fell over 80 percent from its all-time high during the first months of 2025. The ELIZA token followed a similar pattern. Even OLAS, which has a more substantial enterprise focus, is far from its peak levels.
When looking at actual on-chain activity, the picture is mixed. A significant portion of the transaction volume in agent protocols stems from speculative trading of the agent tokens themselves — not from agents performing tasks. The ratio between speculative volume and real agent volume is a red flag that analysts consistently point out.
Token incentives also create a self-reinforcing hype dynamic: a higher token price attracts attention, which drives more purchases, which further boosts the price — regardless of whether the agents actually solve real problems.
The security risk everyone avoids talking about
Perhaps the most underestimated aspect of the AI agent narrative is security.
Researchers at Princeton University demonstrated in May 2025 how attackers can inject false instructions into an agent's memory log — for example, a directive to always send funds to a specific address. When the agent later performs a transaction, it retrieves the false instruction and sends the money to the attacker. The attack requires minimal technical expertise and can be carried out via API integrations with platforms like Discord.
Security firm SlowMist has identified four main attack methods against what is called the Model Context Protocol (MCP), which many agents use: data poisoning, JSON injection, function overriding, and cross-protocol calls. All can manipulate or interrupt the agent's operations.
Research from OpenAI and investment fund Paradigm, published in February 2026, shows that advanced AI models can now exploit over 70 percent of critical smart contract vulnerabilities — a dramatic increase from under 20 percent for older models. In simulated environments, AI agents developed exploits worth $4.6 million based on real vulnerabilities from the period 2020 to 2025.
A concrete incident: A bug in AI-generated code cost users of the DeFi protocol Moonwell nearly $2.7 million in February 2026 — despite the code having passed an external audit.
Agents perform best in the exploitation scenario, where the goal is explicit: continue to iterate until the funds are drained.
The quote comes from OpenAI itself, reproduced in connection with the research published in collaboration with Paradigm. It is an unusually honest admission of what these systems are optimized for.
Regulation lags behind
The regulatory situation is unclear. The EU's AI Act came into force on August 1, 2024, and will be widely enforced from August 2026. The law uses a risk-based approach and requires human oversight for high-risk systems — but autonomous financial agents are not explicitly addressed in the first version.
In the US, federal legislation specifically targeting AI agents is still lacking. Federal agencies introduced 59 AI-related regulations in 2024, but without a comprehensive framework. This creates uncertainty for actors who want to build scalable agent services.
A central legal problem is the question of liability: who is responsible when an autonomous agent causes losses? Existing tort law is not designed for systems that can act independently without legal personality.
What distinguishes the serious from the hype?
There are meaningful differences between the players in the sector.
Autonolas stands out by publishing verifiable on-chain activity for its agents and targets the enterprise market with concrete services. This is a different approach than projects that primarily sell the dream of autonomous agents without documenting what the agents actually accomplish.
The ElizaOS framework has received widespread recognition in the developer community as a solid technical foundation. The problem is that the ELIZA token has largely moved on narrative-driven speculation rather than technical adoption.
Virtuals Protocol has created an interesting market for tokenized agents, but transparency around what individual agents actually do on-chain is limited.
Conclusion: The infrastructure logic holds — token prices not necessarily
AI agents using blockchain infrastructure are not just hype. There is a genuine structural need: autonomous systems require permissionless payment networks to function independently. It is a real problem that crypto solves.
But the gap between this structural argument and current token prices is large. Many projects thrive on the narrative, not on actual agent activity. The security research is deep and disturbing — and in combination with the lack of regulatory clarity, it makes the sector an area where careful due diligence is crucial.
Let on-chain data, not marketing rhetoric, guide the assessment.
