Skip to content

Response, New York. Noah Park: Check Point Patches Actively Exploited Management Server Path Traversal. Detection, Dallas. Ava Okello: Report Finds Nearly Half of Deployed Detections Need Attention Before SOCs Can Trust Them.

SOCtember

Always First. Fast SOC News.

Terminal screenshot from Huntress showing a Python probe sending three MS-SNTP requests to a domain controller for RID 1000, with results legacy 68B reply, extended current-key TIMEOUT and extended old-key 120B reply, followed by one hash line beginning sntp-ms-sha512.
Image: Huntress (fair use). The three-way probe against a live domain controller, Huntress's Figure 3.

Detection

Huntress Reverses the Timeroasting Format Modern Windows Sends by Default, Builds a Cracker for It, and Says the Alert That Holds Is a Host Asking a Domain Controller to Sign for an Account That Is Not Its Own

A Huntress researcher pulled the key derivation for the 120-byte MS-SNTP authenticator out of netlogon.dll, proved it against a patched Windows Server 2025 domain controller and wrote a hashcat module for it. He found packet size a poor signal, and Microsoft Defender for Identity raised no alert across eleven test runs.

Ava Okello, DetectionLondon6 min read

LONDON - Timeroasting, the trick of asking a Windows domain controller to sign a network time reply with a computer account's password and then cracking that signature offline, has a second form that public tools largely ignore and that modern Windows already uses for ordinary time sync, according to research published Nov. 19 by the security firm Huntress.

Andrew Schwartz, the Huntress researcher, wrote that the only detection most teams can copy today is "a volumetric counter of connections to UDP/123 that understands nothing about the protocol," and that the newer form "has no coverage at all." He reverse engineered the Windows binaries that implement it, confirmed the result against a fully patched Windows Server 2025 domain controller and published both a cracker and a set of detection rules.

Two formats, one secret

Timeroasting was described by Tom Tervoort of Secura in 2023. It abuses MS-SNTP, Microsoft's authentication extension to the Network Time Protocol: a client names an account by its relative identifier, or RID, and the domain controller answers with a message authentication code keyed on that account's password, with no authentication required. Sweep the RIDs, collect the codes, crack them offline. Because the targeted accounts never log on, Schwartz noted, there are no 4768, 4769, 4624 or 4776 events for logon-based analytics to evaluate.

The familiar request is 68 bytes and uses MD5, the form hashcat cracks as mode 31300 and that timeroast.py and NetExec produce. The other is 120 bytes. Microsoft added it to the specification as the preferred option because of weaknesses in MD5, Schwartz wrote, and modern Windows sends it by default, to the point that it broke time synchronization against Samba domain controllers. Its exact byte encoding is not fully documented, and hashcat has no module for it.

Reading netlogon.dll, Schwartz found that the time service hands the work to Netlogon, which derives a 64-byte key from the account's NT hash with an SP800-108 counter-mode function using HMAC-SHA512, the 8-byte label "sntp-ms" with its trailing null and the RID in little-endian order, then signs the 48-byte NTP response with a second HMAC-SHA512 pass. He listed four details that "silently" break a cracker if guessed wrong, including the null in the label and the mixed byte order. A Python reference implementation reproduced the domain controller's checksum exactly.

On his test domain controller, the legacy form answered for both the current and the previous key, and the extended form answered only for the previous key, a pattern he reproduced across nine machine accounts. He flagged the current-key timeout as an open observation he could not fully explain.

A cracker, and who is exposed

Schwartz also wrote a hashcat module and OpenCL kernel for the extended form that pass hashcat's own self-test, numbered 31350 as a placeholder until maintainers assign a mode. It benchmarked at 17.6 MH/s on an Apple M3 Pro, he wrote, and is much slower per guess than the MD5 form. But both forms key off the same NT hash, so a weak password falls to either and a random machine secret to neither.

Terminal screenshot from Huntress of a NetExec timeroast sweep against a lab domain controller, listing sntp-ms hash lines for RID 1000 and a cluster of RIDs from 1166 to 1195.
Image: Huntress (fair use).A NetExec sweep in which only computer and trust account RIDs answer, Huntress's Figure 7.

The gaps in a sweep are the point, he wrote. A domain controller signs only for computer and trust accounts, so the realistic targets are the narrow set with guessable secrets: machines built from images or never rotated, hand-set trusts and non-Windows joins. He recommended keeping automatic machine-account password rotation on, fixing predictable pre-Windows 2000 computer-account passwords, limiting access to UDP 123 on domain controllers and, where Seamless SSO is enabled, rolling the AZUREADSSOACC Kerberos key at least every 30 days as Microsoft advises.

What to alert on

Schwartz's main detection finding cuts against intuition. On an older fleet a 120-byte request looked rare; on a modern one, Windows members send 120 bytes as normal time sync, and the public tools, which emit only the 68-byte form, are the odd ones out. "Do not build on packet size," he wrote.

The signal he said holds for both sizes, even for a single packet, is the requested RID compared with the sender. A domain member normally asks the domain controller to sign for its own computer account, so resolving the source IP to an account through DHCP or DNS and then to its RID, and alerting when the two do not match, catches a slow one-RID-at-a-time sweep that no volume threshold would. Domain controllers need separate handling, because they can legitimately use a trust account's RID during hierarchy time sync. He wrote that Suricata cannot do this mapping, so it belongs in Zeek or a SIEM, and that the stock Zeek NTP parser does not expose the field, so a small Spicy analyzer or raw payload access is needed to read the RID at offset 48.

Terminal screenshot from Huntress of Suricata run against a captured timeroast sweep, showing 2 rules successfully loaded, 0 rules failed, 41 alerts for sid 5200002 and 3 alerts for sid 5200010, with the messages MS-SNTP ExtendedAuthenticator USE_OLDKEY_VERSION and MS-SNTP authenticated-NTP sweep to DC (Timeroast, either form).
Image: Huntress (fair use).The two Suricata rules Schwartz would ship, firing on a captured sweep, Huntress's Figure 8.

As a fallback he published a size-agnostic Suricata rule, sid 5200010, that flags 25 client-mode NTP requests from one source to a domain controller within 60 seconds when each is larger than a plain 48-byte request, catching either authenticator size, and an old-key flag rule, sid 5200002, that he said should be treated as a hypothesis until a team has checked whether its own members ever set the flag. His first four rules had failed to load because of inline comments, which Suricata does not support, a mistake he said the engine caught when he tested the rules in Suricata 8.0.7.

He also warned against alerting on a zeroed MAC in the request, which every legitimate client sends by design. Both public timeroast tools he tested send a constant or all-zero transmit timestamp, while a real client's changes every request, a cross-packet anomaly he called a supporting tell.

Terminal screenshot from Huntress of tshark comparing captured NTP requests, showing three timeroast.py requests with byte0=db and the same transmit timestamp and two chrony requests with byte0=23 and changing transmit timestamps.
Image: Huntress (fair use).A timeroast.py request compared with a real chrony client, Huntress's Figure 9.

The identity alert that names it

Microsoft Defender for Identity raised no alert on the traffic across eleven runs over five hours, Schwartz wrote, even though a control DCSync in the same environment drew a high-severity alert within roughly an hour. Microsoft does document an alert, Suspicious SAM Account Name Change, whose description mentions timeroasting. Schwartz wrote that it fires on a sAMAccountName change, the footprint of a targeted variant in which an attacker with write access temporarily disguises a user as a computer account, and does not see the unauthenticated sweep, which touches no directory object.

About the Author

Ava Okello

Ava Okello covers detection engineering, EDR telemetry, and SOC hunting for SOCtember from London.

Related stories

Detection desk