Entry 004 Progresslaya-guard: a free check on every command your agent runs
exit 0Ran itThe regex layer works. 247 tests passed, and all four attack commands were caught.
- Day 3Active
- Day 7Active
- Day 15Upcoming
- Day 30Upcoming
- Day 90Upcoming
- Day 180Upcoming
- Day 365Upcoming
Latest check Day 7, Tue 2026.10.20
- Commits since launch
- 31
- Stars
- 31
- Latest release
- v0.5.2
- Last commit
- 2026.10.19
Numbers by checkpoint
| Checkpoint | Checked | State | Commits since launch | Stars | Releases | Contributors | Last commit |
|---|---|---|---|---|---|---|---|
| Day 1 | Wed 2026.10.14 | ||||||
| Day 3 | Fri 2026.10.16 | Active | 18 | 27 | 9 v0.4.1 | 5 | 2026.10.15 |
| Day 7 | Tue 2026.10.20 | Active | 31 | 31 | 12 v0.5.2 | 5 | 2026.10.19 |
| Day 15 | Due Wed 2026.10.28 | Upcoming | |||||
| Day 30 | Due Thu 2026.11.12 | Upcoming | |||||
| Day 90 | Due Mon 2027.01.11 | Upcoming | |||||
| Day 180 | Due Sun 2027.04.11 | Upcoming | |||||
| Day 365 | Due Wed 2027.10.13 | Upcoming | |||||
What changed
Day 7 Tue 2026.10.20
- Shell commands are judged against the protected paths too. Before, a command that read a key file through `python -c` passed (v0.5.1).
- The protected paths also apply in each agent's own OS sandbox. When a script reads `~/.ssh`, the kernel refuses it (v0.5.2).
- Calls from Claude Code's file tools are judged against the protected paths. The protected paths are credential stores, `.env` files, private keys, shell startup files and `.git/hooks` (v0.5.0).
- An evaluation now replays the commands behind seven public incidents with AI coding agents. It found a delete hidden in a GraphQL call and Windows commands that were blocked without need, and both were fixed (v0.5.2).
- An optional log records the files that the agents' tool calls name. `laya-guard footprint` sums it up per project and suggests paths to protect (v0.5.0).
Day 3 Fri 2026.10.16
- Commands no longer pass when the server is down. If it cannot judge a command, it asks the user or blocks it (v0.4.0).
- Shell comments are removed before the verdict. This blocks attacks that use a comment to lower the verdict (v0.2.6).
- A local dashboard was added. One command installs the hook in all eleven supported agents (v0.3.0).
- The server can run as a service that starts at login and restarts after a crash. An outage shows on the dashboard and raises an alert (v0.4.0).
- A Simulation view was added to the dashboard. It runs eight attack scenarios through the actual evaluator (v0.4.1).