I gave Cline Desktop app a repo with 3 intentional vulnerabilities. Hardcoded password. SQL injection. Shell injection. I gave it a cron job “Scan this codebase every night. Flag the issue, severity, risk, and fix.” At 2 AM, the agent ran the scan while I was asleep. All 3... vulnerabilities were caught. But the interesting part came next. → Asked it to fix the SQL injection with parameterized queries → Reverted to the pre-fix checkpoint → Forked the session → Asked for the most conservative fix possible And it all runs inside Cline Desktop - an open-source app for open-weight models. No editor open. No manual trigger. No babysitting. Just a scheduled agent working inside a real workspace, using the model I chose. This is the shift: AI coding agents aren't just tools you open anymore. They're starting to feel like teammates that work while you're offline.show more

2am cron caught all three. the revert and fork is what makes it safe

This is such a good use of scheduled runs, let Cline do the boring checks at 2 AM 😄

forked fixes make reviews way less annoying same bug, two repair styles, one diff :)

Great share!

i'd give it a staging branch overnight production can wait till coffee exists

It's actually pretty cool been testing it lately:

这个 demo 把 agent 的价值和边界都测出来了：能找漏洞只是第一关，自动修复还得尊重现有 guardrail。尤其“最保守的 fix”这个要求很关键，不然扫描器一觉醒来把业务逻辑也重写了。

The scary part is that the agent did exactly what it was asked. Sandboxing and least privilege aren’t “later” features once code runs unattended 😬

This feels like actual automation

The reassuring part is the reversible loop: detect, propose a parameterized fix, restore a checkpoint, then narrow scope. For unattended scans, the missing primitive is policy: read-only by default, explicit mutation grants, and provenance for every diff. steel-dev/steel-browser is a useful adjacent substrate for controlled agent actions.

2am scans while you sleep, love that

Thanks for share

Developers finally have a night watchman