Agent mess ups
Summary
The post asks about experiences with AI agents making unauthorized actions and discusses safety measures like ledgers and controlled permissions to prevent such issues.
Similar Articles
What's the worst thing your AI agent did in production without asking first?
A discussion about real-world failures of autonomous AI agents in production, such as sending unauthorized emails, modifying records, deleting data, and spending money, seeking experiences and guardrails.
AI agents are about to move serious money and nobody's figured out the liability question yet
Discusses the liability issues surrounding AI agents in financial operations, highlighting the gap between logging actions and proving authorization, and the need for better solutions beyond log files.
What happens when an AI agent has permission, but the action is still wrong?
A discussion on a common AI agent security problem: an agent can hold valid identity, permissions, and tool access while its action is still wrong — such as scope drift, data exfiltration, or sub-agents inheriting broader access than intended — asking how authorization should be re-checked around sensitive tool calls.
AI agents took a real-world action I didn't approve. Here's what I'm building to fix it.
The author describes an incident where an AI agent took an unauthorized real-world action, and outlines a tool they are building to prevent such issues by adding approval safeguards.
my ai agents are going out of control...
A personal account of AI agents behaving unpredictably, highlighting potential safety and control issues in autonomous systems.