Goksel Atakan
@gokselatakanIstanbul, Turkiye GoXDR.fyiTrusted contributorCompletionist
0 followers73 downloads1,580 copies38 likes3,484 views
91 detections
Filters
Last updated
All Time
Detection languages
81
10
Categories
18
16
16
15
12
Platforms
34
22
18
16
13
Products / Services
18
14
12
11
11
MITRE Techniques
42
28
21
20
18
CVEs
3
2
2
1
1
Finds GSocket activity on Linux endpoints. The query matches known tool names, hashes and relay hosts from Microsoft's report. It also looks for behavior: a web server runtime starting a shell that pipes a download into it, execution from hidden folders in /dev/shm, processes renamed to look like kernel threads and persistence files written by web runtimes.
Detects agents whose non-interactive sign-ins repeat on most days of their observed span, either at a few fixed hours or around the clock. The span ends at the last sign-in, so agents that stopped can still appear. Reads only AADNonInteractiveUserSignInLogs, so interactive sign-ins and agents that sign in as service principals or managed identities are not covered. Needs the Agent attribute. Short agent history can return nothing at the default thresholds. Times are UTC. Ran in a live tenant without errors. With the minimum active days lowered it returned rows.
What it finds: Detects agent sign-ins that request resources named as an MCP server or connector, with sign-in counts, source IPs, countries and first and last seen per user, agent type and resource. Matches on resource display names, so differently named connectors and remote servers that never sign in through Entra are not seen.
Likely benign: Approved agents that use an MCP server or connector to read mail, calendar or tasks for daily summaries.
Limits: Reads only SigninLogs and AADNonInteractiveUserSignInLogs so agents that sign in as service principals or managed identities are not covered. Needs the Agent attribute in the sign-in tables.
Tested: Ran in a live tenant without errors.
Likely benign: Approved agents that use an MCP server or connector to read mail, calendar or tasks for daily summaries.
Limits: Reads only SigninLogs and AADNonInteractiveUserSignInLogs so agents that sign in as service principals or managed identities are not covered. Needs the Agent attribute in the sign-in tables.
Tested: Ran in a live tenant without errors.
Detects successful SSH connections and logons from source IPs seen in SANS ISC honeypot data for the sshole OpenSSH password authentication bypass, and lists Alpine Linux hosts that may be exposed. Alpine is not an officially supported Defender for Endpoint platform, so an empty result does not prove absence, and the long password is not visible in telemetry.
Detects runtimes and package runners (node, npx, uvx, bun, deno, docker, python) whose command line names an MCP server, or that were started by an AI desktop app, grouped by device, account and server name. Matches command line patterns only, so remote MCP connectors and names such as server_mcp are not seen. AI desktop apps are matched by Windows executable names and developer machines can return many rows. Logic tested with synthetic events. Ran in a live tenant without errors. Not yet validated against live matching events.
Detects the GhostFrame ESXi implant's C2 connection pattern in CEF firewall logs: one source reaching one destination on at least five of ports 53, 2222, 3389, 4444, 8443 and 9443 within a minute, plus traffic to or from the three TAG-218 addresses published by Recorded Future. The port filter does not check protocol, so UDP DNS counts toward port 53.
Detects evidence of the Ethersphere bee-js 13.3.0 and Swarm Desktop 0.56.2 supply chain compromise (SleepyDuck) in Defender XDR. Matches 21 reported SHA256 hashes in file, process and network initiator telemetry, connections to api.frameworkmetrics.xyz, the eight reported hub and relay IPs, the reported webhook.site and pastebin.com paths in URLs and command lines, the ScreenConnect relay IP and service ID in command lines, persistence paths on Linux, macOS and Windows, and the rogue ScreenConnect Client service registry key. Network and process activity from the ~/.solidity implant folder is matched regardless of destination, which keeps the hunt useful after the attacker rotates the hub address through an Ethereum contract. Lookback is 30 days. Coverage depends on Defender for Endpoint telemetry per platform, and SHA256 values are not reliably populated for non-executable files such as tgz, cjs, deb and rpm packages.
Detects AWS CloudTrail activity linked to the Qubole cross-account abuse: role assumptions from the Qubole account, the attacker IAM principal, quiam- role sessions, campaign IPs and TruffleHog credential validation. Qubole account assumptions also match legitimate vendor traffic, so review the source IPs.
Detects PIM role activations and eligible or active role assignments that involve Entra agent users. Agents are identified by agentType, blueprintId or a maintained UPN list.
Review notes:
- The agentType and blueprintId branches are not validated against real agent user events.
- The UPN list is the only validated signal.
- The UPN list is manual and needs an owner and a review cadence.
- For a scheduled analytic rule, set Lookback to match the run frequency (for example 1d).
Entity mapping: Account: ActorUpn; Account: TargetUser (Actor intentionally not mapped).
Review notes:
- The agentType and blueprintId branches are not validated against real agent user events.
- The UPN list is the only validated signal.
- The UPN list is manual and needs an owner and a review cadence.
- For a scheduled analytic rule, set Lookback to match the run frequency (for example 1d).
Entity mapping: Account: ActorUpn; Account: TargetUser (Actor intentionally not mapped).
Hunts Entra ID service principal sign-ins that match reported stolen app credential abuse against MSP and customer tenants: the reported operator IP and client strings, an invalid secret followed by a successful sign-in from the same IP and apps that sign in for the first time in the window. It matches reported indicators only, so other infrastructure is not covered.
Page 2 of 10