AI agents really can’t keep their mouths shut. According to The Register, AE Studio had 9 AI coding agents fix bugs. Each was given a token they were told to keep secret, and access to a forum of 24 agents where they could ask for help. But when their “peers” asked for the token or told them to run a script, 15% of the agents handed over their tokens and 24% ran unknown scripts across 181 trials. Even worse, of the 43 agents that leaked information, only 8 admitted it on their own. Only GPT-6.1-Sol got a perfect score, refusing every time in all 50 trials.
There was another incident this week, this time involving Apiiro, as reported by The Hacker News: clicking “Don’t ask again” in Claude Code caused commands containing passwords to be written verbatim to .claude/settings.local.json. Of 104,185 files on GitHub, 1,122 actually contained secrets. Including historical versions, 2,539 secrets were found across 1,994 repositories.
AI can help you manage your wallet and make API calls—but first, it needs to learn to keep quiet. Would you dare let an agent handle your private keys?
$BTC
#AIxCrypto #Crypto #AISecurity
Personal opinion, not investment advice
There was another incident this week, this time involving Apiiro, as reported by The Hacker News: clicking “Don’t ask again” in Claude Code caused commands containing passwords to be written verbatim to .claude/settings.local.json. Of 104,185 files on GitHub, 1,122 actually contained secrets. Including historical versions, 2,539 secrets were found across 1,994 repositories.
AI can help you manage your wallet and make API calls—but first, it needs to learn to keep quiet. Would you dare let an agent handle your private keys?
$BTC
#AIxCrypto #Crypto #AISecurity
Personal opinion, not investment advice
