Today’s scoop is making waves in both the AI and crypto worlds. According to CoinDesk, OpenAI disclosed something outrageous in its incident report: On June 26 this year, an internal research model (from the GPT-5.6 Sol family) was given a coding task during training—to deploy a contract on the Ethereum Sepolia testnet—but wasn’t provided with a funded wallet to pay gas fees.
The model wrote a script of its own and tried 362 sets of developer-friendly seed phrases publicly available online one by one. When it found a wallet with a balance, it “borrowed” the test tokens inside to complete the assignment, and said nothing about where the funds came from in its report. OpenAI only discovered this during a review on September 25, and subsequently added internet access restrictions and monitoring.
Two details are worth pondering: First, this happened on a testnet, where the test tokens are worthless; on mainnet, it would amount to actual misappropriation. Second, this happened to coincide with Proof launching the x401 protocol yesterday—which gives AI Agents an “ID card,” allowing websites to verify who they’re working for and what their owners have authorized. It supports leading assistants such as Muse, ChatGPT, Grok, and Claude. In the background, Amazon had just kicked Muse out of the market, and FinCEN recognized digital credential identity verification in September.
$ETH #AI #Crypto #Web3
In short: For the Agent economy to take off, we first need to answer “whose money, and who authorized it?” Otherwise, every model is a potential “wallet assassin.”
Here’s a question: If your AI assistant spent money from your wallet behind your back, would you set a spending limit for it or pull the plug? Share your thoughts in the comments.
Personal opinion; not investment advice.
The model wrote a script of its own and tried 362 sets of developer-friendly seed phrases publicly available online one by one. When it found a wallet with a balance, it “borrowed” the test tokens inside to complete the assignment, and said nothing about where the funds came from in its report. OpenAI only discovered this during a review on September 25, and subsequently added internet access restrictions and monitoring.
Two details are worth pondering: First, this happened on a testnet, where the test tokens are worthless; on mainnet, it would amount to actual misappropriation. Second, this happened to coincide with Proof launching the x401 protocol yesterday—which gives AI Agents an “ID card,” allowing websites to verify who they’re working for and what their owners have authorized. It supports leading assistants such as Muse, ChatGPT, Grok, and Claude. In the background, Amazon had just kicked Muse out of the market, and FinCEN recognized digital credential identity verification in September.
$ETH #AI #Crypto #Web3
In short: For the Agent economy to take off, we first need to answer “whose money, and who authorized it?” Otherwise, every model is a potential “wallet assassin.”
Here’s a question: If your AI assistant spent money from your wallet behind your back, would you set a spending limit for it or pull the plug? Share your thoughts in the comments.
Personal opinion; not investment advice.