At 06:29:21 UTC on 4 August, someone who had been inside since a quarter to four in the morning opened msconfig.exe and forced a reboot. The machine came back in Safe Mode with Networking, a boot in which Windows does not load third-party services. The EDR agent stayed out. Defender logged an error of its own, 0x8007043c: "this service cannot be started in Safe Mode".
The case was published by Huntress on 12 August, and one thing is worth saying up front: the agent that went dark was theirs. They wrote it up anyway, with the event IDs and the timestamps. We run managed EDR/MDR on other companies' estates and we staff the on-call rota, so the report reads from here as a list of what we should have caught.
The timeline, with the times there are and the ones there aren't
Everything below is in UTC, which is how Huntress publishes it. Two phases of the attack carry no timestamp in the report, and we are not going to invent one here either.
- 03:45A burst of failed logins against several usernames on a SonicWall SSL VPN, from several external IPs. A textbook credential spray.
- 03:52:42A valid account logs in from
72.23.77[.]35. Roughly seven minutes after the first attempt, with no second factor in the way. - +2 h"Nearly two hours" of silence, the report says, and then RDP into the domain controller: an elevated
cmd.exe, a ping and a full directory dump withGet-ADUserandGet-ADComputerintoAdUsers.txtandAdComp.txt. No timestamp published. - no timeA pivot to the application server, WinRAR installed mid-intrusion to archive the mapped file shares, and an upload to an attacker S3 bucket with
s5cmd. No timestamp either. - 06:29:21
msconfig.exeand a forced reboot. The boot writesKernel-Boot27 withLoadOptions = NOEXECUTE=OPTOUT SAFEBOOT:NETWORKandKernel-General12 withBootMode = 2. Shortly before that — the report doesn't say how long — they had registered AnyDesk under theSafeBoot\Networkkey so they would not be locked out too. - 06:34:29
akira.exestarts. At 06:36:33 the system logs event 26, "Out of Virtual Memory", and a second later PowerShell dies because it cannot create a new guard page for the stack. - 07:43:50A scheduled Defender scan identifies the binary as
Ransom:Win32/Akira.B!ibt. Cleanup fails over and over with0x80004005: there is detection and no remediation, because real-time protection is still dead. - 08:10:38The attacker reboots the host back to normal operation. At 08:12:28, with real-time protection loaded again, the file is quarantined. The person who put it there is the one who got it removed.
One hour forty-one, not ten
Some of the coverage puts the EDR-less window at ten minutes. That number is in the report, but it measures something else: Huntress says the AnyDesk session, including the preceding Safe Mode reboot, lasted less than 10 minutes. The blind window is a different sentence, two paragraphs earlier: for the entire Safe Mode window, the host had no working EDR, and AV was blinded.
That window opens at 06:29:21 and closes at 08:10:38. It is one hour, forty-one minutes and seventeen seconds with the machine on the network, answering, with no agent and no real-time antivirus. The subtraction is ours, on the timestamps Huntress published, because the report doesn't state the duration.
What stopped the encryption was memory
Safe Mode boots with a stripped-down environment and constrained virtual memory. Huntress phrases it carefully and we are going to borrow their caution: the Akira process tree appears to have starved it, and the "Out of Virtual Memory" box lines up exactly with the moment the payload tried to get going. Their sentence on this goes in whole, because the second half is the one that matters: "a lucky side effect of the attacker's own mistake in these circumstances, not a defence you can plan around".
And they add the detail that finishes the argument: a host with more physical memory or a larger page file might have given akira.exe enough virtual memory to encrypt in Safe Mode. The same intrusion, on a server with 256 GB, ends somewhere else. So the honest summary of this incident contains the word luck, and that word never describes the incident: it describes the security programme underneath it.
The files left before the reboot
The report doesn't date the exfiltration, but it does order it: WinRAR over the mapped file shares and an upload with s5cmd before the 06:29:21 reboot, which in turn comes before the 06:34:29 encryptor. By the time the attacker touches the binary, the files are already in a bucket that isn't yours. Textbook double extortion: the encryption is there to hurry along someone who has already lost the data.
Worth saying even though it makes us look worse: a flawless backup would have changed nothing about that day. It gives you the files back and leaves the blackmail lever untouched. We sell managed backup and even so the right order in this post is access first, watching second, and the copy as the safety net underneath. And if your backup server lives inside the same domain they have just dumped wholesale with Get-ADComputer, that is a different problem we wrote up separately.
In Safe Mode your agent is not essential to Windows
There was no exploit, no signed vulnerable driver, no kernel tricks. There was an msconfig.exe and a reboot, and the operating system's own design did the rest: the minimal driver and service set excludes third-party security products by definition. MITRE has a technique for it, T1688, and Huntress points out that families such as Snatch and AvosLocker have done it for years. What they are seeing for the first time is Akira doing it.
It fits what the quarter looks like as a whole. In its Q2 2026 report, Halcyon counted 1,988 publicly claimed attacks and 89 active groups, with Akira fourth by volume on 119. Infosecurity opened its piece on that report by summing it up this way: shutting endpoint detection tools down before encryption begins has become standard operating procedure across the ransomware ecosystem.
The line almost nobody quoted
It sits in the mitigations section and reads fast: in this environment, the agent was on a fraction of the machines the attacker enumerated. That is where an attack gets prepared, on the hosts nobody watches, which is why the Get-ADComputer at the start isn't a forensic curiosity: it is the map that tells the attacker which ones. An EDR rollout with holes turns your inventory into somebody else's.
This part lands closest to home. When we take over someone else's estate, the first sum we do in managed EDR/MDR is how many hosts are in the directory and how many carry an agent. The gap between those two numbers is the attack that hasn't happened yet.
The times are UTC, and that changes the conversation
Huntress doesn't say where the victim was, so converting to local time is our exercise and not a fact of the case. Had this happened in mainland Spain, the spray falls at 05:45 and the valid login at 05:52 — small hours, nobody about — and the blind window would run from 08:29 to 10:10. Mid-morning. With the office open.
We prefer that scenario to the easy headline, because it is more uncomfortable. A machine that reboots on its own and goes without an agent for an hour and three quarters can slip past at ten in the morning on a Tuesday, with everyone at their desk, and hiring a night shift doesn't fix that. What fixes it is 24x7 support wired to the console, and a rule saying that a silent agent wakes somebody up, whatever the hour.
Seven minutes at the door
Everything above happens afterwards. The door was an SSL VPN with a username and a password, and it opened in roughly seven minutes of spraying. It doesn't matter how good the rest is if the perimeter gives to a reused password, and we have written about this in another guise: when the thing that fails is the perimeter appliance itself, and when there are authentication paths that never get around to asking for the second factor.
The mitigations, in the order Huntress puts them
The list is theirs and so is the order. In brackets is what we do with each point, which is the only thing here that is ours:
- 1Catch the spray before it becomes a foothold. Alert on bursts of failures against several usernames from one source, and correlate them with a successful login from the same IP or ASN in a short window. (We build this on the VPN concentrator's own logs, which are almost always there with nobody reading them.)
- 2Shut the door. MFA on every VPN account, an IP allowlist while an attack is live, and rotation of VPN and domain credentials after a compromise, treating everything in the
Get-ADUserdump as exposed. (The "this service account can't do MFA" exception is the one that ends up becoming the rule.) - 3Close the coverage gaps. EDR on every host, and VPN plus Windows logs into a SIEM: the first VPN logons were visible hours before detonation. (That margin of hours only exists if someone receives the alert.)
- 4Watch for the Safe Mode play. Boot-configuration changes (
msconfig.exe,bcdedit),Kernel-Boot27 with aSAFEBOOTload option,Kernel-General12 withBootMode=2, security services stopping (System event 7036) and tooling added to the minimal-service key. (Of all of them, the one that has served us most often is the agent that goes quiet.)
What we are NOT going to tell you
- ✗To change your EDR. The agent that was left out of that boot was a good one, and in Safe Mode any other third-party agent would have vanished as well. Switching brand here only changes the name of the one that was missing.
- ✗That this is fixed by buying something. Three of the four mitigations are configuration of what you already own and of logs that are already being written. What costs money is someone receiving the alert.
- ✗That this attack failed. They took the file shares and the full map of the directory. The encryption collapsing saved work for the people restoring, and none at all for the people negotiating.
In short
A VPN with no second factor, a dump of the entire directory, file shares archived with WinRAR and pushed to S3, and a reboot that left the host one hour and forty-one minutes without an agent or an antivirus. The encryptor was brought down by Safe Mode's virtual memory, and Huntress warns that on a host with more RAM it wouldn't have been. Everything else in the case — events 27, 12, 26 and 7036, the burst against the VPN, the agent going quiet — was being written into logs that most environments already generate.
Sources: timeline, timestamps, commands (Get-ADUser, Get-ADComputer, WinRAR, s5cmd, reg.exe add on SafeBoot\Network, msconfig.exe), events Kernel-Boot 27, Kernel-General 12, System 26 and Defender 3002 / 0x8007043c, the Ransom:Win32/Akira.B!ibt detection with cleanup failing 0x80004005, their own agent being the one disabled, technique T1688 with the Snatch and AvosLocker precedents, the partial EDR coverage and the mitigations in their published order, plus the quote about the lucky side effect — Huntress, 12 August 2026, read in the original rather than in summaries. Coverage placing the EDR-less window at ten minutes — BleepingComputer; reproduction of the side-effect quote — SC Media. Quarterly figures (1,988 attacks, 89 groups, Akira fourth on 119) and the opening summary about EDR shutdown — Halcyon's Q2 2026 report, as covered by Infosecurity Magazine. Our own arithmetic: the 1 h 41 min 17 s window (06:29:21 → 08:10:38) and the conversion to mainland Spanish time, neither of which Huntress publishes.
How many hosts are in your directory, and how many carry an agent?
At everyWAN we deploy and watch managed EDR/MDR, with 24x7 support behind it so the alert reaches a person. We review agent coverage, the MFA on your remote access and what happens when a host stops reporting — and we also tell you when you already had it right.
Talk to everyWAN