Eight minutes. Elastic says that is, in some cases, the average time between an infostealer landing on a customer's machine and an attacker logging in somewhere with the credentials it took. That leaves defenders very little time to react.
The figure comes from the November 2026 edition of "threats, etc," a quarterly report from Elastic Security Threat Command, published Nov. 4, 2026, by Terrance DeJesus, Jason Pappalexis and Andrew Pease. It draws on confirmed customer intrusions and Elastic's own research from April through September 2026. Elastic's write-up of the report is titled "When trust becomes the payload," which sums up its argument. Across the cases it describes, attackers mostly didn't break things. They borrowed things defenders already trust: valid logins, signed remote-management software and, in a couple of cases, AI coding and chat assistants that were already installed on the machine.
The eight-minute window
According to Help Net Security's coverage, the report counts 27 confirmed infostealer cases on customer computers. Infostealers are malware built to grab saved browser passwords, session cookies and similar secrets and send them back to the operator. In those 27 cases the average gap between infection and an attacker using the same credentials to log in was eight minutes.
That number matters because a lot of incident response still assumes there is time. A familiar playbook goes: detect the malware, isolate the host, open a ticket, then get around to rotating the user's passwords. If the stolen credentials are being used within minutes, the account compromise has already happened by the time anyone rotates them, and the endpoint is no longer the whole story.
Elastic also notes that adversary-in-the-middle (AiTM) phishing kits and consent loopholes are being used to get around multifactor authentication. AiTM kits sit between the victim and the real login page and capture the session after MFA succeeds. Consent abuse tricks a user into granting a malicious app access to their account. Neither one needs the attacker to defeat MFA directly.
ClickFix keeps working
Help Net reports that ClickFix-style pages were the most common way in, accounting for 12 of the 27 infostealer cases. ClickFix lures are fake error messages, CAPTCHA checks or "fix this problem" prompts that talk the user into copying a command and running it on their own machine. The user does the attacker's work, so defenses focused on scanning email attachments can miss it.
The report shows how fast this goes once someone takes the bait. In one healthcare case, Elastic says the intrusion went from a paste-and-run lure to delivery of an MSI installer that used DLL sideloading in roughly 16 seconds. (DLL sideloading means getting a legitimate, often signed, program to load a malicious library placed next to it.) IT Brief Australia describes the same case as malware arriving about 16 seconds after the user responded to the lure. A human analyst is unlikely to intervene in a 16-second window, which favors prevention and automated containment.
Signed remote-access tools as the backdoor
The report records 23 confirmed cases of remote monitoring and management (RMM) abuse, with ScreenConnect named most often. RMM tools are what managed service providers and internal IT teams use to support machines remotely. They are signed and widely deployed, which makes them attractive to attackers: Elastic's point is that people trust them because IT likely installed them and they carry a trusted signature.
This isn't a ScreenConnect vulnerability story. The point is that a legitimate, signed binary set up to phone home to an attacker-controlled instance gives the attacker persistent, interactive access that looks a lot like routine support.
When the attacker drives your AI assistant
The AI tools section is the headline item. Help Net describes two intrusions:
- Claude Code on Windows. An attacker who had connected to the machine remotely used the Claude Code extension for VS Code to prepare tools and reach an outside server.
- ChatGPT desktop app on macOS. The attacker used the app to read an internal infrastructure document and then accessed the machine's saved password store.
Help Net is clear about what this shows and what it doesn't. In its words: "These cases involved misuse after attackers obtained access. They do not establish that the applications contained a vulnerability or that their providers were compromised." The attackers were already inside, and they acted through assistants that were already installed. Elastic's report frames AI applications as points for command execution rather than a new type of attack.
Elastic also describes a separate intrusion into a Kubernetes cluster that involved an AI agent. Elastic reconstructed about 17,600 actions over four days. IT Brief puts the pace at roughly one action every 20 seconds. Indicators that an agent may have been involved include duplicate commands issued seconds apart and a coding error immediately after a successful exploit, which fits automated activity. Elastic doesn't attribute the activity to any named group, and the sources here don't either.
Help Net's piece also quotes Mike Nichols, Elastic's general manager of security, on the report's findings.
The common thread
Read together, the cases make one point. Each technique abuses something the defender has already approved. Credentials are valid until they're revoked. RMM tools are signed and expected. AI assistants are installed on purpose and given broad local access because that's what makes them useful. ClickFix gets the legitimate user to run the code. Controls that ask "is this binary malicious?" or "is this login using the right password?" will pass all of it.
What It Means for Defenders
Who's affected: any organization with users who browse the web, store passwords in browsers, run RMM software or have deployed AI coding and chat assistants on endpoints. Elastic's healthcare example is a reminder that regulated sectors are not exempt. The cases described here involve abuse of trusted software and credentials rather than a vulnerability, so the fixes are largely in configuration, monitoring and response speed.
- Treat an infostealer detection as a credential compromise. With an average of eight minutes before the stolen logins are used, revoke active sessions and rotate passwords and tokens for the affected user as part of the first response step, not after cleanup. Check sign-in logs for that identity from the time of infection onward.
- Plan for MFA being bypassed. Elastic points to AiTM phishing kits and consent loopholes as working MFA bypasses. Phishing-resistant MFA such as FIDO2 security keys or passkeys holds up far better against AiTM relays than push or SMS codes. Restrict which third-party apps users can grant consent to, and review existing app grants.
- Close the ClickFix gap. Teach users that a legitimate website will not ask them to copy and paste a command into a Run dialog or terminal. Where your platform supports it, limit or alert on users launching command interpreters and on script engines spawning installers, and make sure detections act automatically. The healthcare case moved in about 16 seconds.
- Inventory and allowlist RMM tools. Know which remote-access products your organization actually uses and which tenants or instances they connect to. Alert on, or block, any RMM tool or instance that isn't on that list. ScreenConnect was the most frequently abused tool in Elastic's 23 cases, so if you don't use it, its appearance deserves immediate investigation.
- Bring AI assistants into scope. Treat agents like Claude Code and desktop chat apps as privileged software: log what they run, limit what files, password stores and network destinations they can reach where possible, and include their activity in hunting. Watch for the agent-like patterns Elastic describes, such as near-duplicate commands seconds apart and rapid, high-volume action sequences.
- CMMC and federal contractors: these cases align with access control, identification and authentication, and audit logging practices. This is our inference, not a point the sources make: unapproved RMM tools and unmonitored AI agents on systems that handle controlled unclassified information could raise assessment concerns.