Endpoint case · built on 2026 attacks
The coding assistant read an SSH key
An AI coding assistant ran a command on a file in the .ssh folder. Doing what the developer asked, or what a tool told it to?
- AdvancedLevel
- 35 minAbout
- 3 versionsEvery attempt
- KQL or SPLQuery language
Free. Read the briefing as a guest; investigating needs a free account.
The alert
AI coding agent ran a command on a file in the .ssh folder
The AI coding assistant on RID-WS-0517 ran PowerShell that read something in jjohnson's .ssh folder. These agents run commands all day, so it might be fine. I do not know how to tell.
From one version of the case: the company, the people and the details change with every attempt.
What you do
- Read the escalation from an L1 colleague and the events they attached.
- Hunt the logs in KQL (Microsoft Sentinel) or SPL (Splunk) and pin the evidence, then prove it: answer questions with values from the logs.
- Find the root cause and scope the incident: which accounts and hosts are affected.
- Map what you found to MITRE ATT&CK, set the severity and choose the response.
- Get a debrief: what happened, the evidence you missed, and what the attacker does next if you were wrong.
Each attempt is another version: another company, other people, and sometimes another outcome. One version may be an attack, another a false alarm. You cannot learn the answer, only the method.
Logs you will query
| Microsoft Sentinel | Splunk |
|---|---|
Syslog | index=linux sourcetype=linux_secure |
Sysmon | index=sysmon sourcetype=XmlWinEventLog |
WebProxy_CL | index=proxy sourcetype=cisco:cloud_security:proxy |
What it is built on
2026 research on hijacking AI coding agents through their tools: malicious MCP servers, and injected data read by legitimate ones (Agentjacking used forged Sentry error reports). The agent in the case is fictional. The companies, people, domains and addresses are fictional; the behaviour and the log formats follow the public reporting.


