Agent systems should be designed assuming prompt-injection and exfiltration attempts. Separating harness and compute helps keep credentials out of environments where model-generated code executes.

Youtube Kids is an example of how the product designed for kids differs from the one targeting adults. It’s much easier to navigate thanks to bigger buttons and fewer content boxes on the page. Plus the security settings on the platform make sure that younger users are safe and have access to appropriate content. Those all are parts of a thought-through design interface for children.

Do not pass arguments right into subshell, it's as unsafe as eval.

npx link is a tool I developed as a safer and more predictable alternative to npm link.

Strings generated by devalue can be safely deserialized with eval or new Function

The NYCLU found nothing in the documents outlining policies for accessing data collected by the cameras, or what faces would be fed to the system in the first place. And based on emails acquired through the same FOIL request, the NYCLU noted, Lockport administrators appeared to have a poor grasp on how to manage access to internal servers, student files, and passwords for programs and email accounts. “The serious lack of familiarity with cybersecurity displayed in the email correspondence we received and complete absence of common sense redactions of sensitive private information speaks volumes about the district’s lack of preparation to safely store and collect biometric data on the students, parents and teachers who pass through its schools every day,” an editor’s note to the NYCLU’s statement on the Lockport documents reads.