Nº 03
Sandboxes, gateways and the limits of agent security.
Built-in agent sandboxes, isolated runtimes, gateways, classifiers and EDR each draw a line in a different place. What each one governs, and what walks around it.
Read entry
In this issue
03 / 03
Nº 02
Humans are leaving the loop. Now the boundary has to be real.
On August 14, Auto Mode becomes the default in Claude Code: a model reviews each command instead of a human approving it. It caught 89% of dangerous commands in testing; humans caught 13.6%. That is a real improvement, and it is still not a security boundary.
Read entryNº 01
Three coding-agent incidents, three different failures.
In the past two months, three coding-agent incidents have surfaced publicly: a Cursor agent deleting a production database, a Gemini agent exfiltrating a token after prompt injection, and an npm supply-chain compromise that hit any host running npm install. Each one isolates a different structural property of how agents fail, and the defenses don’t overlap.
More to come
Relm Journal