Use case · On our own infrastructure
Patch a critical CVE in minutes, without handing an agent SSH keys or a root password
On our own servers, an AI agent found the exposure and applied the fix. Alpacon held it to one scoped work session and recorded its commands.
- from the session's first approval to Apache serving patched PHP
- ~6min
- The agent never held an SSH key or root password
- No keys
- to assess CVE-2026-98432, a CVSS 9.8 kernel CVE, and put a mitigation in place, about 15 minutes in all
- 3sessions
With Alpacon: a scoped work session
- Request a work session: servers, job, time limit
- A person approves
- Each command is checked against the session
- It runs on the in-scope servers
- Commands in the session are recorded
What happened: a PHP remote-code-execution CVE
0 min
Scoped work sessionOne server, one hour
The agent wrote down the job. Someone on the team approved it.
1 min
Not in scope yetsudo failed
The session didn't include sudo yet, so the command failed.
3 min
Approved by a personsudo for a named list
The agent asked for nine command patterns. Someone on the team approved them.
5 min
Checked against scopePHP patched
18 packages upgraded, Apache restarted on patched PHP.
10–13 min
Commands recordedOrigin locked down
The record shows why each sudo command was allowed.
Who did what, minute by minute
Scroll to see the full timeline →
CVE-2026-98432 is a use-after-free in the Linux kernel's IPv4 TCP handshake code, rated CVSS 9.8 in its CVE record.
Scroll to see the full timeline →
The kernel CVE took three scoped sessions: an exposure check, a check for signs of compromise, then a recheck that put a mitigation in place. No fixed kernel had shipped for the server yet, so as a precaution the agent disabled MPTCP with a persistent sysctl setting; the kernel itself stays unpatched until a fixed kernel ships. About 15 minutes in all.
Our agents' work sessions since June 2026
We run our own infrastructure on Alpacon
- work sessions our agents opened
- 457
- commands recorded
- 4,675
- commands refused before they ran
- 78
- median session: access for one job, then it ends
- 2.3 min
FAQ
Does the agent get root access?
What if the agent tries something outside the job?
Which agents work?
Playbook: critical CVE response
See it on your servers
- Exposure check
- Scoped fix
- Approval for sudo
- Full record
On our own infrastructure, October 2026. Times from the Alpacon session record.