Agent Tesla v4 hides in emoji and never hits disk: an email-rule problem, not a new-malware one
KnowBe4 says a new build of the commodity infostealer uses Unicode emoji to break string signatures and keeps its final payload off disk. The lure is a spoofed bank thread aimed at finance teams.
The Watch · Security desk
What happened
- KnowBe4 research identified a new version of Agent Tesla, version 4, an infostealer observed being dropped via a business email compromise lure targeting finance departments.
- The email observed by KnowBe4 arrived as a forwarded thread presenting as internal correspondence forwarded to an account's contact.
- The attackers spoofed the address of Metropolitan Bank and Trust Company, a legitimate Philippines-based commercial bank.
- The email thread was designed to look like an in-progress discussion the recipient had been brought into late, and the recipient was directly instructed to confirm an attached document and reply.
- The malware operates via a JScript dropper that can be launched with a simple initial open-with dialog.
Why it matters
KnowBe4 researchers logged an attempted delivery of Agent Tesla version 4 via a business email compromise lure aimed at finance departments, and published their analysis on August 20 [1][24]. Nothing about the payload's ambitions is new; what has changed is that the delivery chain breaks string-based signature matching and file-based scanning at the same time, which leaves the mail gateway and script-level rules as the enforcement points [25].
The lure did the unglamorous work. According to KnowBe4, the message arrived as a forwarded thread presenting as internal correspondence passed to a contact of the account [21], with the address of Metropolitan Bank and Trust Company, a Philippines-based commercial bank, spoofed as the sender [22]. The thread was built to look like an in-progress discussion the recipient had been brought into late, with a direct instruction to confirm the attached document and reply [23]. That is indistinguishable from a normal accounts-payable Tuesday.
The attachment is a JScript dropper that can be launched from a simple open-with dialog [2]. Its body is interleaved with Unicode emoji characters, hearts and water droplets among them [3], which disrupt string-based signature matching and make the code visually noisy enough to defeat casual review [4]. On execution the script writes two files to C:\Users\Public\Libraries\, one of which is misdirection, and the extension is passed into DonutLoader shellcode for reflective PE injection [5]. The final Agent Tesla binary never touches the filesystem, so file-based scanners have nothing to open [6].
The payload itself is a catalogue of the usual defences: obfuscation with ConfuserEx [7], embedded metadata presenting the assembly as a Python installer [8], and a standard Windows debugger check that halts execution if it finds a debugger attached [9]. Before collection it builds a persistent hardware fingerprint so operators can track a victim across reinstalls and IP rotation [10], and it disables validation for all outgoing connections to keep C2 traffic from raising alerts or errors [11]. Collection covers browsers, messaging platforms and native Windows credential repositories [12] across more than 40 applications [13], plus keystroke and clipboard capture [14]. Every exfiltrated file carries a header with timestamp, username, computer name, OS, CPU, RAM, public IP and the MD5 hardware ID [15].
The operational detail that should shape your playbook is the timing. KnowBe4 reports the credential dump lands on the attacker's FTP server within seconds of execution, with no delayed staging [16], sent to a single threat actor-controlled domain [17]. There is effectively no dwell window between click and credential loss, so post-execution containment is a rotation exercise, not an interception one [26].
The mitigation KnowBe4 proposes is correspondingly cheap. The firm says the emoji obfuscation does not survive any YARA rule looking for the Unicode code points used alongside JScript-specific patterns [18], and states that "a rule matching both the emoji distribution pattern and WScript.Shell or CreateObject calls will catch this family" [19]. Their advice to defenders is to update email security rules rather than wait for the endpoint to catch it [20].
Two things worth checking this week. First, whether your mail filtering inspects script attachments at the code-point level at all, since that is the specific gap the technique exploits [18]. Second, whether finance staff have any procedure that treats being forwarded into a live thread late, with an instruction to confirm an attachment, as a reason to verify out of band [23].
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence45
- Adoption20
- Hype gap+30
- Incentives70
- Confidence52
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
KnowBe4 research identified a new version of Agent Tesla, version 4, an infostealer observed being dropped via a business email compromise lure targeting finance departments.
- [2]
The malware operates via a JScript dropper that can be launched with a simple initial open-with dialog.
- [3]
The script body contains a number of Unicode emoji characters, such as hearts and water droplets, interleaved directly through the code.
- [4]
The emoji characters obfuscate the malicious file by disrupting string-based signature matching and making the code visually noisy enough to defeat casual review.
- [5]
Once launched, the script writes two files to C:\Users\Public\Libraries\, one of which is a misdirection, with the extension passed into DonutLoader shellcode for reflective portable executable injection.
- [6]
The final Agent Tesla binary never touches the filesystem and cannot be detected by file-based scanners.
- [7]
Agent Tesla v4 is intentionally scrambled using an obfuscator tool called ConfuserEx to make it nearly unreadable to analysts.
- [8]
The malware's assembly presents itself as a Python installer in its embedded metadata.
- [9]
The malware uses a standard Windows function as its first line of defence to detect whether it is being watched by a debugger, and stops running if it finds one.
- [10]
Before harvesting credentials the malware creates a persistent hardware fingerprint, allowing attackers to track victims consistently across reinstalls or IP rotations.
- [11]
The malware disables validation for all outgoing connections, maintaining communication with C2 infrastructure without triggering security alerts or errors.
- [12]
The malware sweeps credentials from web browsers, messaging platforms and native Windows credential repositories.
- [13]
The final payload is configured to sweep credentials from more than 40 applications.
- [14]
Agent Tesla is capable of intercepting keystrokes using its keylogger and clipboard tool.
- [15]
All exfiltrated files include a system fingerprint header: timestamp, username, computer name, OS name, CPU, RAM, public IP and the MD5 hardware ID.
- [16]
KnowBe4 noted that the credential dump lands on the attacker's FTP server within seconds of execution, with no delayed staging.
- [17]
The payload rapidly exfiltrates harvested credentials via file transfer mechanism to a single threat actor-controlled domain.
- [18]
KnowBe4 said the emoji-obfuscation approach in the JS dropper does not survive any YARA rule that looks for the Unicode code points used alongside JScript-specific patterns.
- [19]
KnowBe4 researchers said: "A rule matching both the emoji distribution pattern and WScript.Shell or CreateObject calls will catch this family."
- [20]
The researchers advised security teams to update their email security rules to catch Agent Tesla before it can harvest credentials.
- [21]
The email observed by KnowBe4 arrived as a forwarded thread presenting as internal correspondence forwarded to an account's contact.
- [22]
The attackers spoofed the address of Metropolitan Bank and Trust Company, a legitimate Philippines-based commercial bank.
- [23]
The email thread was designed to look like an in-progress discussion the recipient had been brought into late, and the recipient was directly instructed to confirm an attached document and reply.
- [24]
The KnowBe4 blog containing the analysis was published on August 20.
- [25]
Because the dropper's emoji padding defeats string-based signature matching and the injected payload never lands on disk for file-based scanners, the remaining detection opportunities in this chain are at the email layer and in script-content rules.
- [26]
With credentials reaching the attacker's FTP server within seconds of execution and no delayed staging, there is effectively no dwell window in which to interrupt exfiltration, so post-execution response is credential rotation rather than interception.
Sources
2 independent publishers whose own reporting we read for this story.
- infosecurity-magazine.comNew Agent Tesla Malware Variant Boosts Evasion Capabilities
1 article · August 21, 2026
- scworld.comNew Agent Tesla malware version uses emoji obfuscation to evade detection
1 article · August 21, 2026
Topics and entities
Topics
- Infostealer Malware
- Malware Obfuscation and Evasion
- Credential Exfiltration
- Business Email Compromise
- Detection Engineering
Entities
- SC World
- YARA
- Metropolitan Bank and Trust Company
- KnowBe4
- Infosecurity Magazine