AI agent Astra patched NixOS servers with a 20-second outage
The author of the "LLM under the hood" Telegram channel tasked an AI agent named Astra with finding vulnerabilities on his servers and applying patches right away. One server was unavailable for about 20 seconds, but the risk was low: on NixOS, a failed system can boot the previous system image. The agent found one vulnerability, CVE-2026-97032, and several minor issues in NPM. The takeaway: it is easier to hand routine admin work to an agent when there is a cheap rollback.
AI-processed from TG: @llm_under_hood (LLM под капотом); edited by Hamidun News
The author of the Telegram channel "LLM under the hood" tasked an AI agent named Astra with checking his servers for fresh vulnerabilities and applying patches right away. While the update was running, one of the servers was unavailable for about 20 seconds. The author had no reason to fear breakage: nearly all of his infrastructure runs on NixOS, where a failed system can boot the previous system image. The agent found one vulnerability with the number CVE-2026-97032 (this is how publicly registered security holes are named) and several minor issues in NPM, that is, in JavaScript packages that had not been updated for a long time.
The note is short, but it shows a working approach: when an AI agent can be given administration, and what can cushion the risk.
What the agent actually did
The trigger was anxiety. The author had heard company stories about hacking attempts, and, in his words, it got to him. He gave the agent a short task in English. He admits he is scared of recent hacking stories (in the prompt he refers to the latest conversations in "brain"). Next, he asked it to check the servers tied to the bitgn_platform and to fleet management centers. It was to find serious vulnerabilities reachable from the internet that should be closed quickly. The output was required to be a list of three priorities.
The agent did not stop at a report. It went through the servers and updated kernels, userland (system programs and libraries on top of the kernel), and individual applications. During this update, one server dropped off the network for about 20 seconds. The author mentions this at the start of the note so readers are not surprised.
The result was calm. The author calls CVE-2026-97032 not scary but unpleasant: the servers can be rebooted remotely. The rest are minor issues in the NPM stack. After the fixes, he asked the agent to automatically repeat the audit in a couple of weeks and write to Telegram if anything new appears.
Why trusting the agent is not frightening here
The main reason is NixOS. In this system, server configuration is described as code, and every state is saved as a separate build. Think of a game with saves before every boss: if an update broke a service, you load the previous system image and return to a working state. The author states directly that this is why trusting the agent does not scare him.
For practice, a simple rule follows. The danger is not the agent with administrator rights itself, but the lack of a cheap rollback. When any change can be undone quickly, the cost of a model's mistake drops sharply. Then it can be given dull, repetitive work: reading vulnerability lists, matching them against installed versions, and applying patches. On ordinary servers with hand-made configuration, the same scheme would require disk snapshots, maintenance windows, and a more cautious mode.
What to take from this
First, the request does not have to be complicated. The original prompt has no long instructions. It has context (I am worried about recent hacks), boundaries (servers reachable from the internet), and an output format (top three priorities). For an audit of this type, that was enough.
Second, the agent is convenient to turn into a regular check. The author did not stop at a one-time cleanup but ordered a repeat audit with Telegram notification. This way, security stops being a crisis after alarming news and becomes a background process.
Third, it is not worth copying the approach without a safety net. The post does not show what permissions the agent had, how its actions were verified, or exactly how many servers were updated, so the scale cannot be compared with corporate infrastructure. Roughly 20 seconds of downtime is acceptable for personal services, but for a payment system or an API with strict availability commitments, such a pause may be unacceptable. The findings are modest too: one CVE and outdated NPM packages. This is more a hygiene check than proof that the agent can catch sophisticated attacks.
The practical conclusion for small teams is this. If you have reproducible builds, rollback, and a clear boundary of permissions, an AI agent can take on regular updates and a first-pass audit. If you do not, start with rollback first, and only then give the agent access to the servers.
- Nov 21, 2026A viral GTA 6 Claude clip looks fake: half an hour of control burned through the 5-hour Pro limit
- Nov 21, 2026Claude Code Speeds Up Snowboard Kids 2 Decompilation: From 20% to 45% of Functions in Three Months
- Nov 21, 2026VibeCBS collects project nicknames in advance: the agent finds 2.8 times more documents
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
Key Questions Answered
What the agent actually did?
The trigger was anxiety. The author had heard company stories about hacking attempts, and, in his words, it got to him. He gave the agent a short task in English. He admits he is scared of recent hacking stories (in the prompt he refers to the latest conversations in "brain"). Next, he asked it to check the servers tied to the bitgn_platform and to fleet management centers. It was to find serious vulnerabilities reachable from the internet that should be closed quickly.
Why trusting the agent is not frightening here?
The main reason is NixOS. In this system, server configuration is described as code, and every state is saved as a separate build. Think of a game with saves before every boss: if an update broke a service, you load the previous system image and return to a working state. The author states directly that this is why trusting the agent does not scare him. For practice, a simple rule follows. The danger is not the agent with administrator rights itself, but the lack of a cheap rollback.
What to take from this?
First, the request does not have to be complicated. The original prompt has no long instructions. It has context (I am worried about recent hacks), boundaries (servers reachable from the internet), and an output format (top three priorities). For an audit of this type, that was enough. Second, the agent is convenient to turn into a regular check. The author did not stop at a one-time cleanup but ordered a repeat audit with Telegram notification.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.