Ethereum’s AI Security Agents and a $2,000 Bitcoin Trap Reveal Crypto’s New Risk Lab
•
October 25, 2026 7:55 pm
•
Comments
Crypto security is starting to look less like a closed-door audit and more like a live laboratory.
Ethereum researchers are now coordinating AI agents to hunt for the second-order effects of proposed protocol changes. At the same time, a deliberately vulnerable website invited the public to steal roughly $2,000 in Bitcoin—and somebody did.
The two experiments are different in scale and purpose, but they point to the same uncomfortable truth: automated defenses and public stress tests can expose weaknesses faster, yet neither removes the need for careful human verification.
AI agents enter Ethereum’s Glamsterdam review
Cointelegraph reports that the second and final testnet rehearsal for Ethereum’s Glamsterdam upgrade was canceled after ChainSecurity found vulnerabilities connected to EIP-8037 and EIP-8038. The proposed gas-scaling changes could alter how contracts behave when a call runs out of gas, with at least 13 DeFi contracts identified as potentially affected.
That is exactly the kind of protocol change that can look tidy in isolation and become dangerous once it touches years of assumptions buried in production code. The Hoodi rehearsal is not expected to return until researchers have mapped those downstream effects more fully.
Derek Chiang of Ethlabs described an unusual part of the response: researchers from Ethlabs and the Ethereum Foundation independently deployed groups of AI agents to scan contracts, then created a GitHub-based message board so the agents could coordinate their findings.
— Derek Chiang | Ethlabs (@decentrek) October 25, 2026
The interesting part is not that an AI system found a magic answer. Multiple automated reviewers can probe a large contract surface in parallel, exchange leads, and give human researchers a sharper list of places to investigate.
A second board has reportedly been opened to examine EIP-8198, the “Quick Slots” proposal aimed at a later upgrade.
A Bitcoin challenge proves the value of hostile testing
The week’s other experiment was more direct. A website called “Steal My Bitcoin” intentionally left a private key reachable through an unpatched PHP environment.
The test was built around the PHP2Shell class of vulnerabilities that prompted organizations in several countries to take systems offline while emergency patches were prepared.
The site drew heavy traffic and probing. A solver ultimately extracted the key, demonstrating that the vulnerable server path was exploitable.
But the result also exposed a very human failure: organizer James Pavur acknowledged that the recovered key was not initially the one tied to the wallet displayed on the site.
Congrats to @killa who solved https://t.co/a4rI7FPsln hours ago but I screwed up which key went on the server (the flag was a root wallet WIF instead of the branch WIF).
Check out their awesome writeup here: https://t.co/vnui8sJ5mj
— James Pavur (@JamesPavur) October 25, 2026
That mistake matters. Security exercises can prove that an exploit works while still producing the wrong operational conclusion if the challenge itself is configured incorrectly.
The exploit was real; the first wallet mapping was not.
The lesson is coordination, not autopilot
Ethereum’s experiment shows where AI-assisted review can be genuinely useful: covering more code paths, comparing assumptions, and surfacing edge cases before a protocol change reaches mainnet. The Bitcoin challenge shows why the output still needs to be checked against the real target.
Neither story supports handing security over to autonomous systems. Both support a more practical model in which machines widen the search, public challenges add adversarial pressure, and humans remain responsible for validating what was actually found.
That combination may be less dramatic than “AI saves crypto,” but it is far more useful. The next generation of security work will probably be faster, more collaborative, and more automated.
It will not be self-verifying.
Join the conversation!
We have no tolerance for comments containing violence, racism, profanity, vulgarity, doxing, or discourteous behavior. If a comment is spam, instead of replying to it please click the icon below and to the right of that comment. Thank you for partnering with us to maintain fruitful conversation.