Kris 🧙‍♂️ Kowal

@kriskowal.com

https://kriskowal.com/

Your system prompt and the attacker's payload share one context window, both competing for the model's attention. You already know that fight goes badly some percentage of the time. Real enforcement has to live somewhere the model can't talk its way past.

Build revocation so you never need the agent to cooperate. Keep the revoke handle separate from the capability the agent carries. Something looks wrong, you pull it from the outside, and the agent learns about it by failing.

Robin and Ramy witness Babel, the Royal Institute of Translation, in all its foreboding majesty second illustration for R. F. Kuang's Babel - illustrated edition, out now from Harper Voyager 40x30cm ink on claybord

a black and white scratchboard drawing. In the background a very large tower looms over Gothic buildings overlooking a vast paved plaza. In the foreground, two characters wearing black Oxford robes walk towards the tower.

npm v12 is now generally available. npm install now makes install scripts, Git, and remote-URL dependencies opt-in by default. We're also retiring npm 2FA-bypass GAT: no account management (early Aug 2026), no direct publishing (~Jan 2027). More info at github.blog/changelog/20...

npm install-time security and GAT bypass2fa deprecation - GitHub Changelog

npm v12 is now generally available and tagged latest. This major release turns on the install-time security defaults we announced in June, and it’s also where we begin a deprecation…

github.blog

"We have full audit logs" is the most reassuring sentence in agent infrastructure, and the most misleading. An audit log catches the second incident, not the first. It is a forensic tool, not a control.

It is an interesting thing that it seems like no matter how you feel about AI use, any communication that you receive that smells of AI often elicits a kind of negative vibe

The industry spent years trying to make LLMs ignore malicious instructions. Turns out the cheaper bet is making sure a compromised agent can't reach anything it wasn't handed. Least authority. References as permission. We've had this since the E language.