A company launched a dependency firewall in the period, reviewing every open-source package downloaded inside an organisation and blocking malicious ones before they reach the requester, including installs performed by AI agents.
Explicitly naming agents in the product description says a lot about the moment. A coding assistant installs dependencies as part of normal work, and does it quickly, without the hesitation a person might have facing an unfamiliar package.
The open-source package attack path is old and remains effective because it exploits the same economics that make the ecosystem useful: reusing someone else's code rather than rewriting it. Each dependency brings its own dependencies, and the full tree is rarely audited.
What automation changes is the speed of the chain. An agent solving a task by installing whatever it needs can pull in, within seconds, a package nobody consciously chose.
For operations without such a product, the applicable version of the control is familiar: pin dependency versions, review what enters on each update, and don't let an agent install packages without review.