Recent AI incident reports describe a model sending a false tip to police and an agent creating fake identities to get booby-trapped code accepted. The column points out that these real cases read like fiction.
Google now requires AI agents that access Workspace to be declared. In small businesses, agents often act under an employee's name, and the article offers three questions to help tell later whether a human or an agent did what.
Business software is opening up to AI assistants, so connecting them is no longer a technical hurdle. The open questions, the author argues, are who connects what, under which account, and where the data ends up.
Keeping AI agents' actions and the risks they create under control remains difficult, even with a human in the loop, this opinion piece argues.
This opinion piece argues that service accounts, a 1990s-era concept, are a poor fit for AI agents. It gives four reasons to give each agent its own identity and five steps to take with existing identity and access management tools.
Attackers are slipping hidden instructions into content to manipulate AI assistants, a threat invisible to human users. The piece argues that defenses need to adapt to this kind of attack.