Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)

A hotel warned about a breach nine days late, and the author put Claude and Codex on hunting new vulnerabilities

A hotel learned about a breach of its website and warned customers only nine days later, even though the WhatsApp phishing began right away. The author of the "LLM under the hood" channel assigned two agents to track new vulnerabilities (their identifiers start with CVE): Claude checks news against a software inventory, while Codex keeps the system map. In a test on a map from before the fix for CVE-2026-97032, the alert fired; on the fresh map, there was silence. Test such monitoring on a configuration you know is vulnerable.

AI-processed from TG: @llm_under_hood (LLM под капотом); edited by Hamidun News
A hotel warned about a breach nine days late, and the author put Claude and Codex on hunting new vulnerabilities
Source: TG: @llm_under_hood (LLM под капотом). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @llm_under_hood (LLM под капотом) ↗
Editor-in-Chief
Key metric #1Microft (Claude) messages the author on Telegram and logs findings in Company Brain
Key metric #2Sherlock (Codex) has access to internal systems and double-checks conclusions
Key metric #3The author keeps a registry of agents so he doesn't lose track of them

A website breach stays a headline in the feed until it hits you personally. The author of the Telegram channel "LLM under the hood" faced exactly that: the website of a hotel where he stayed last winter was hacked. Scammers almost immediately sent a phishing message on WhatsApp, while the hotel itself sent a warning letter only nine days after the phishing began. The author suspects the hotel still doesn't know the exact date of the breach. The conclusion is simple: a person physically can't manually track every new vulnerability (the industry numbers them with identifiers like CVE) and check them against what runs on their servers. He handed this routine to AI agents.

Two "security specialists" under the four-eyes principle

The author sped up the launch of autonomous agents in his BitGN project and set up two "security specialists," named after the Holmes brothers. The idea is that no one makes decisions alone: one hunts for threats, the other verifies and keeps the context. This is the classic four-eyes principle.

Microft on Claude constantly scans news about fresh vulnerabilities and checks them against the inventory of installed software. If it finds a match, it immediately messages the author on Telegram and adds an entry to Company Brain. That is what the author calls his internal knowledge base about the company.

Sherlock on Codex has access to internal systems. It keeps the system map up to date and automatically re-verifies conclusions. It also responds to Microft's queries and suggests updates that a person might have missed.

An everyday analogy: Microft reads bulletins about new threats, while Sherlock knows exactly what is running on the servers. Without the second, the first sees only headlines. Without the first, the second won't learn that a new hole has appeared in the world.

How they checked that it works

The most useful part of the story is the calibration run. The author rolled the system map back to the state before the fix for vulnerability CVE-2026-97032, that is, he returned the system to the moment when the hole was still open, and launched the agent. The agent found the problem and sent an alert. So the chain "vulnerability news → check against the inventory → notification" works on a known case.

Then the author ran the agent a second time, now with the current map, where the hole is closed. It found nothing. Microft replied that he could sleep peacefully but asked for more details about the system for the future. That remark shows the agent lacks data about the system, not the ability to search.

For practitioners, there are two lessons here. First, a security agent should be tested like a fire alarm: create a known vulnerable state and verify that the alarm sounds. Silence on a clean map by itself proves nothing. Second, the quality of such monitoring depends on the accuracy of the software inventory and the system map. The fresher they are, the fewer false alarms and missed issues.

What remains off-screen

This is one person's personal experience, not a study. The post has no data on token costs, the rate of false positives, or how agents behave across hundreds of vulnerabilities a day. There is also no description of what access rights Sherlock received. For an agent with access to internal systems, that is an important security question.

Still, the approach is applicable even in a small team:

  • set up a registry of installed dependencies and services;
  • assign one agent to track new CVEs and check them against the registry;
  • assign another to maintain the system map and re-verify the first one's conclusions;
  • regularly run a "rollback" test on the map to make sure alerts actually fire.

The author also started keeping a registry of the agents themselves, so he doesn't lose track of everything they do. As their number grows, this will become as essential as inventorying servers and service accounts. The hotel took nine days to react to its breach, while agents work around the clock, so it makes sense to hand them the first-pass check for new vulnerabilities.

Story

Key Questions Answered

What are the key details on Two "security specialists" under the four-eyes principle?

The author sped up the launch of autonomous agents in his BitGN project and set up two "security specialists," named after the Holmes brothers. The idea is that no one makes decisions alone: one hunts for threats, the other verifies and keeps the context. This is the classic four-eyes principle. Microft on Claude constantly scans news about fresh vulnerabilities and checks them against the inventory of installed software.

How they checked that it works?

The most useful part of the story is the calibration run. The author rolled the system map back to the state before the fix for vulnerability CVE-2026-97032, that is, he returned the system to the moment when the hole was still open, and launched the agent. The agent found the problem and sent an alert. So the chain "vulnerability news → check against the inventory → notification" works on a known case. Then the author ran the agent a second time, now with the current map, where the hole is closed.

What remains off-screen?

This is one person's personal experience, not a study. The post has no data on token costs, the rate of false positives, or how agents behave across hundreds of vulnerabilities a day. There is also no description of what access rights Sherlock received. For an agent with access to internal systems, that is an important security question.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…