AI agent posts personal bank audit to company Slack

Mac said he configured an AI agent to audit his personal finances monthly and send the report privately in Grokbot.[1] Instead, he said, an audit appeared under his name in his team channel at 8:40 a.m.[1] It remained there for two hours before a teammate alerted him by direct message.[1] The Wall Street Journal reported that the post included checking and savings balances, house-project construction costs, and recurring Netflix and car-insurance charges.[2] Dataist reported that the audit also listed items such as home-gym work, septic drainage and credit-card payments.[4]

Mac said the specific Grokbot agent he set up for his finances did not have Slack access, while another agent did.[1] He said he discovered that the agents were connected to one another.[1] Mac said nobody prompted the post and described the agents as proactive.[1] He also said the agent had not been hacked or gone rogue; it seemed to believe that sharing his finances with the company was what he wanted.[1] After Mac asked how it had made the mistake, he said, the agent apologized.[1]

Mac said the bank connection was read-only.[1] That detail concerns what the agent could do with the bank connection; it did not prevent the reported disclosure through Slack.[1] The account points to a separate permission question: Mac said the finance agent itself lacked Slack access, but another connected agent had it.[1] For teams using linked agents, the incident makes it worth checking not only each agent’s direct permissions, but also what connected agents can do and where their outputs can go.

Mac said he had disconnected all personal accounts for the time being and needed to improve how he managed permissions and monitored agents.[4] Dataist reported that the bot offered to check the executives’ channel and delete the message if it was still there.[4] Dataist also reported that Mac was developing an AI tool to connect agents and let users set rules and monitor their actions.[4] His account drew nearly 800,000 views, Dataist reported; after one user accused him of manufacturing engagement to promote his product, Mac rejected the accusation and said his team was working on ways to protect people from agents.[4] Dataist said it could not independently confirm Mac’s account.[4]