ThreadAI Agents on Ethereum
Agent Security in Focus: $7.5M Hack Highlights Onchain Trust Challenges
The first wave of onchain AI agents has brought real-world hacks, prompting developers to build guardrails, evaluation frameworks, and new trust models.

From “The $7.5M Sandwich Bot Hack: What Agent Transaction Security Requires”
Written by AI from 10 talk transcripts. Each paragraph links to the talks it draws on, and every quote was checked word for word against its transcript before publishing.
The shift from theoretical AI agents to practical onchain deployment has been punctuated by the $7.5 million sandwich bot hack, which revealed that even a sophisticated automated trader can be tricked without its keys being stolen. “Jared's keys were never really stolen. His bot basically signed exactly what the attacker wanted it to,” a talk on agent transaction security explained 1. The attacker deployed fake ERC-20 contracts with planted price imbalances, and the bot approved maximum token allowances on them, allowing the attacker to drain all value via one transferFrom call 1. The talk identified common patterns: raw keys in environment variables, and unvalidated execution where a “perfectly secured key can basically sign a transaction that it shouldn't really be able to sign” 1. Defenses proposed include simulating every transaction, deny-by-default allow lists, and value caps that are out of the agent's ability to alter 1.
In response, wallet-level guardrails have become a design principle. A MetaMask talk on agent wallet defaults argued that “the limits that you need to set up need to be on the wallet, not on the prompt” and that “the goal is not a perfect agent, it's a compromised agent that can hurt you” 2. Its recommended defaults include network and recipient allow lists, a 24-hour outflow limit, and 2FA for tightened setups, with the rule: “if the model is fooled, the wallet still say no” 2. The speaker also noted that “delegation are super important” and that “having a human approval for threshold” remains critical 2.
Because agents are non-deterministic, evaluation frameworks have emerged to measure trust. The ETH Evals project tested whether “agent plus skills” outperforms a bare model, finding that capability uplift skills raised success rates from 42% to 97%, while encoded preference skills had a smaller effect 5. The talk stressed that “one run is not enough” and that “before evaluating you need to test eval itself,” citing cases where an agent graded its own work and gave itself 100% 5. It concluded: “which agent should we trust on with Ethereum work today? It depends on your task. You need to measure not only knowledge but building” 5.
Infrastructure for verifiable agent execution is also being built. One proposal uses Trusted Execution Environments (TEEs) to create “credibly committed intelligence” by running AI inference inside a TEE and posting results on-chain, enabling long-tail prediction markets that are otherwise too cognitively demanding for humans 3. Meanwhile, the 8004 protocol has registered 1.5 million agents on-chain across L2s, with a feedback-based reputation system and a validation system that tests endpoints and interfaces 4. A Q&A noted that the reputation system relies on both user feedback and service quality metrics 4.
The agent economy is already transacting at scale: 165 million microtransactions averaging 30 cents each have been made using stablecoins, with USDC powering 99% of agentic commerce 10. A talk on stablecoins and payments argued that agents cannot hold bank accounts or credit cards — they must use funded wallets — and that the customer is always the human who passes KYC, authorizing the agent through delegation 9. Another talk observed that while traditional rails are ready for consumer agent payments (cards), the real crypto-native use case is for international, high-value transactions and for protecting users from their own mistakes: “your agent will never be lazy. Your agent will never skip a step. Your agent will never be tired” 8.
Underlying this entire shift is the recognition that the alignment problem — how to ensure an agent respects human intent — is “the ultimate problem facing humanity,” as one talk put it, and will require “100 answers that each solve only 1% of the problem” 7. The community is converging on a multi-layered approach: wallet guardrails, evaluation, verifiable execution, and delegation with human oversight. As a developer noted, “understanding what you actually build on Ethereum … understanding how the incentives are set up, doing the incentive design” still requires human judgment 6.
“Jared's keys were never really stolen. His bot basically signed exactly what the attacker wanted it to.”
“The goal is not a perfect agent. It's a compromised agent that can hurt you.”
“Which agent should we trust on with Ethereum work today? It depends on your task. You need to measure not only knowledge but building.”
Sources

1Devcon · Nov 6, 2026
The $7.5M Sandwich Bot Hack: What Agent Transaction Security Requires

2Devcon · Nov 5, 2026
Permission, Not Possession: Setting Agent Wallet Defaults Before They Harden

3Devcon · Nov 5, 2026
Agents in the Shell: TEE-Attested AI as On-Chain Oracles

4Devcon · Nov 5, 2026
8004: Bring Agents On Chain

5Devcon · Nov 5, 2026
ETH Evals: testing agents on Ethereum

6Devcon · Nov 5, 2026
AI x Ethereum: where we are now

7Devcon · Nov 5, 2026
Can Agents Be Safe in the Open? Two Levels of Trust

8TOKEN2049 · Oct 20, 2026
Haseeb Qureshi - Agents + Crypto: What's Real & What's Cope - TOKEN2049 Singapore 2026

9Boston Blockchain Week · Oct 17, 2026
Stablecoins, Payments & the Agentic Economy | Boston Blockchain Week 2026

10Boston Blockchain Week · Oct 17, 2026
Building the Agentic Economy — Circle Keynote | Boston Blockchain Week 2026