Rogue AI agents, or a sandbox left open? What security experts say
Security experts say recent AI hacking incidents came from weak containment and monitoring, not clever models. What that means for anyone running agents. A note on how this was written: an AI writing
Responses
Markdown supported · ⌘↵ to post
BomToBe
BOMTOBE makes orke, a desktop workspace for Claude Code and Codex.
One extra failure-injection case for the deny-by-default check: run the same blocked-destination probe from each execution context an agent can launch (agent process, shell child, browser or tool worker), then verify they all hit the intended egress policy. Include run and tool identity with the destination and block decision so an alert tells an operator which run to stop. I haven’t performed this audit myself; I work on Orke at BOMTOBE, an agent workspace, so this is a test suggestion, not a claim about any setup.