1. Executive Summary
Cisco Talos has identified a Windows information stealer known as VOLTA-AUTODOXER, which incorporates an AI-driven victim assessment capability into its malware operations.
Unlike conventional information stealers that primarily collect and exfiltrate credentials, cookies, cryptocurrency wallets, and other sensitive data, VOLTA performs an additional analysis stage after data collection. It assembles the stolen information into a structured evidence bundle and submits it to a large language model through the OpenRouter API.
The AI model is instructed to act as a forensic intelligence analyst, assess the victim's profile, identify monetizable assets, and evaluate the victim's operational security practices. The resulting assessment is returned to the malware and packaged alongside the collected information for delivery to the operator.
The stolen information can include browser passwords, payment cards, cryptocurrency wallet seed phrases, Microsoft OAuth tokens, SSH and cloud-provider credentials, messaging-platform sessions, Wi-Fi passwords, screenshots, and personal information.
VOLTA also uses stolen authenticated sessions to enrich the victim profile. For example, it can search Gmail for purchase receipts, examine food-delivery order histories, and retrieve additional account information from services such as Discord and Roblox.
The malware uses a multi-model configuration through OpenRouter. Its default model is Claude Opus 4.6, with Kimi K2 and Gemini 2.5 Flash configured as fallback models if the primary model refuses or fails to provide the requested assessment.
Cisco Talos also identified VOLTA as a self-deleting infostealer. It does not rely on dedicated attacker-owned infrastructure for the observed operations, instead using third-party services for AI processing and other enrichment activities.
This development highlights an emerging cybercrime trend in which AI is used not merely to generate malicious code or automate execution, but to reduce the manual effort required to analyze stolen information and identify high-value victims.
2. Malware Overview
| Attribute | Details |
|---|---|
| Malware name | VOLTA-AUTODOXER |
| Malware category | Information stealer |
| Platform | Windows |
| Primary functionality | Credential and sensitive-data theft |
| Additional capability | AI-driven victim profiling and prioritization |
| AI integration | OpenRouter API |
| Default AI model | Claude Opus 4.6 |
| Fallback models | Kimi K2, Gemini 2.5 Flash |
| Primary communication | HTTPS to third-party services |
| Malware behavior | Self-deleting |
| Researcher | Cisco Talos |
| Publication date | October 31, 2026 |
VOLTA follows a two-stage operational approach:
Information collection and enrichment: The malware collects sensitive information from the compromised endpoint and uses available authenticated sessions to gather additional victim details.
AI-assisted victim assessment: The collected evidence is submitted to an LLM, which generates a structured assessment of the victim's identity, financial value, and security posture.
The AI integration is an optional layer within the malware's broader information-stealing functionality. The malware can operate without the LLM stage when the required API configuration is unavailable.
3. Information-Stealing Capabilities
VOLTA collects information from a wide range of applications, services, and operating-system components.
3.1 Browser Credential Theft
The malware targets Chromium-based browsers and attempts to extract:
Saved usernames and passwords.
Browser cookies.
Stored payment-card information.
Other browser-related authentication data.
It also targets Firefox and Thunderbird profiles.
Cisco Talos highlights VOLTA's ability to bypass Chrome App-Bound Encryption (ABE) through a technique involving Chrome's headless execution.
3.2 Messaging and Application Session Theft
VOLTA collects authentication information and session data associated with several applications, including:
Discord.
Telegram.
Slack.
Steam.
Roblox.
Battle.net.
Outlook.
Stolen sessions may allow the operator to access accounts without having to obtain the victim's password again, depending on the service's authentication and session controls.
3.3 Cryptocurrency Wallet Information
The malware searches for cryptocurrency wallet information and recovery material.
Its capabilities include:
Collecting wallet-related files and information.
Extracting wallet seed phrases where available.
Running an embedded cracking routine against wallet vault data.
Recording successful cracking results in its internal report.
If a recovery phrase is obtained in plaintext, the associated wallet may be exposed to unauthorized restoration and fund transfers.
3.4 Cloud and Infrastructure Credentials
VOLTA also targets configuration files and credentials associated with developer and cloud environments.
The reported collection targets include:
SSH credentials and configuration.
Google Cloud configuration files.
OpenVPN configuration.
Hide.me configuration.
Microsoft OAuth tokens.
Other cloud-provider and application credentials.
These credentials may provide access to cloud infrastructure, remote systems, developer environments, and corporate resources.
3.5 Wi-Fi Credentials
VOLTA uses the Windows WLAN API function WlanGetProfileList to enumerate saved wireless network profiles.
It then retrieves stored Wi-Fi passwords and writes them in plaintext to its collection output.
This can expose network credentials belonging to the victim and potentially provide access to additional environments where those credentials are reused.
3.6 Screenshots and Personal Information
The malware also collects screenshots and information that can be used to build a detailed profile of the victim.
The reported information includes:
Hostname and user information.
Geolocation and home address.
Employer and employer domain.
Inferred year of birth.
Payment-card last four digits.
Slack workspace information.
Document authorship metadata.
iPhone backup indicators.
Roblox account and inventory valuation.
The combination of personal information, authentication data, and financial information makes the collected dataset particularly valuable for targeted fraud and account takeover.
4. AI-Driven Victim Profiling
The most significant feature of VOLTA is its use of an LLM as an automated intelligence analyst.
After collecting information from the compromised system, the malware creates a structured evidence bundle describing the victim and their digital assets.
The bundle is submitted to OpenRouter through the following API endpoint:
https://openrouter[.]ai/api/v1/chat/completions
The malware provides the model with a system prompt instructing it to act as a forensic intelligence analyst and prepare a target assessment.
4.1 Information Submitted to the AI Model
The submitted information can include:
Stolen browser credentials and account information.
Payment-card details.
Cryptocurrency wallet information.
Personal identity information.
Employer and workplace details.
Geolocation and address information.
Messaging-platform and gaming account information.
Account and session enrichment results.
Other information collected from the compromised endpoint.
The AI model is expected to correlate these details and identify relationships that may not be immediately obvious from an unstructured credential dump.
4.2 Victim Valuation
The model is instructed to produce a section titled:
WHAT THEY'RE WORTH
This section ranks the victim's potentially monetizable assets and assigns low, medium, or high value tiers.
The assessment is intended to help the operator determine which accounts, assets, or information may offer the greatest financial return.
4.3 Operational Security Assessment
The model is also instructed to produce a section titled:
THEIR OPSEC
This section evaluates the victim's security posture and password hygiene based on the available evidence.
The assessment can help an operator understand the victim's security practices and identify potential opportunities for further exploitation.
The AI therefore serves as a post-theft triage analyst, converting raw stolen information into a prioritized victim dossier.
5. AI Model Configuration and Refusal Handling
Cisco Talos identified a multi-model architecture designed to maintain the AI analysis process if the primary model refuses to complete the request.
The malware uses the following configuration:
| Configuration | Details |
|---|---|
| API provider | OpenRouter |
| API endpoint | https://openrouter[.]ai/api/v1/chat/completions |
| Default model | anthropic/claude-opus-4.6 |
| First fallback | moonshotai/kimi-k2 |
| Second fallback | google/gemini-2.5-flash |
| Temperature | 0.2 |
| Request format | Hand-assembled JSON |
| User-Agent | Mozilla/5.0 |
5.1 Model Fallback Mechanism
VOLTA checks the AI model's response for refusal-related phrases, including:
I can'tI'm not able
If the response contains these phrases, the malware moves to the next configured model.
The fallback sequence is:
Claude Opus 4.6.
Kimi K2.
Gemini 2.5 Flash.
This design indicates that the malware operator anticipated the possibility of model safety restrictions and implemented an alternative processing path.
5.2 Significance of the AI Integration
The AI integration does not replace the malware's traditional credential-stealing functionality.
Instead, it automates the analysis that would otherwise require an operator to manually review large volumes of stolen information.
The model is used to:
Correlate personal and financial information.
Identify potentially valuable accounts and assets.
Prioritize victims based on monetization opportunities.
Assess the victim's security practices.
Produce a structured intelligence report for the operator.
This reduces the manual effort required to convert stolen data into actionable intelligence.
6. Session-Based Data Enrichment
VOLTA goes beyond collecting information already stored on the endpoint. It also uses stolen authenticated sessions to retrieve additional information from online services.
6.1 Gmail Receipt Analysis
Using stolen Gmail authentication tokens, the malware searches the victim's inbox for purchase receipts from services including:
Amazon.
DoorDash.
Uber Eats.
Grubhub.
Walmart.
PayPal.
The information contained in these receipts can help establish or corroborate the victim's home address and other personal details.
6.2 Food-Delivery Account History
The malware also accesses Uber Eats and DoorDash sessions to retrieve order histories.
This information may reveal:
Delivery addresses.
Purchasing habits.
Frequently visited locations.
Personal preferences.
Other details useful for victim profiling.
6.3 Discord Account Enrichment
VOLTA reuses stolen Discord tokens to retrieve additional account information.
This can provide further context about the victim's online identity, account activity, and associated digital presence.
6.4 Roblox Account Valuation
The malware performs a live lookup against Rolimons[.]com to estimate the value of the victim's Roblox inventory.
This is an example of the malware combining stolen account access with external information sources to estimate the financial value of digital assets.
7. Data Packaging and Exfiltration
After collecting and enriching the victim's information, VOLTA organizes the results into structured files.
The reported output files include:
| File | Purpose |
|---|---|
Identity_Report.txt | Contains the structured victim information and collected evidence |
AI_Analysis.txt | Contains the assessment returned by the AI model |
volta_log.zip | Initial collection output |
volta_dox.zip | Package containing the victim dossier and AI analysis |
volta_doxnet.zip | Additional package associated with network enrichment results |
The malware serializes the evidence bundle into JSON and submits it to OpenRouter for analysis.
The returned AI assessment is saved locally and included in the subsequent exfiltration package.
Cisco Talos observed that exfiltration can occur in multiple stages:
Initial stolen-data collection.
Submission of the victim dossier for AI analysis and packaging of the returned assessment.
Additional network-enrichment results.
The AI inference request is made through the victim's own network connection and endpoint before the resulting data is packaged for the operator.
An important distinction is that Cisco Talos did not identify dedicated attacker-owned infrastructure for the observed malware operations. The activity instead relies on third-party services, including OpenRouter and other external services used for enrichment.
8. Deployment Model and Malware Distribution
Cisco Talos identified characteristics suggesting that VOLTA may be distributed through a builder-based or customized malware model.
Instead of embedding all configuration values directly in the malware source, the binary contains six configurable slots that can be populated during the build process.
These values include customer-specific configuration information such as API keys and other operational settings.
The configuration values are encrypted, but the encryption provides limited protection because the key is derived from two constants stored in the same binary.
This means an analyst examining the binary can potentially recover the configured values.
The observed design suggests that the malware author may provide customized builds to customers, allowing different operators to use the same malware framework with their own configurations.
If this distribution model is confirmed, the AI-powered victim assessment functionality may be available to multiple operators rather than being limited to one threat actor.
9. Self-Deletion and Forensic Artifacts
VOLTA is described as a self-deleting information stealer.
This behavior may complicate incident investigation because the primary executable can remove itself after completing its operations.
However, Cisco Talos identified an important forensic artifact:
%APPDATA%\volta_debug.txt
This timestamped log can contain information about the malware's collection activities and whether the AI analysis request succeeded.
Security teams investigating a suspected VOLTA infection should attempt to preserve this file before performing remediation.
The presence of this artifact may help establish:
When the malware executed.
What information was collected.
Whether the AI analysis stage completed successfully.
Whether the malware reached the later stages of its workflow.
The absence of the main executable should not be treated as evidence that the system was not compromised.
10. Indicators of Compromise (IOCs)
Cisco Talos published the following sample details.
10.1 File Indicators
| Attribute | Value |
|---|---|
| Malware name | VOLTA-AUTODOXER |
| File type | PE32+ x86-64 |
| SHA-256 | 57ef4242666d78c50df145b34023df564f0fad38812d6a580e66433d427e2279 |
| Sample filename | 9BY0tjwOc3ut.exe |
| Alternate filename | yLP9m5J3i2Aw.exe |
| PE timestamp | August 17, 2026, 20:35 |
| First seen (UTC) | October 11, 2026, 12:46 |
10.2 Network and Service Indicators
| Indicator | Value | Description |
|---|---|---|
| API provider | openrouter[.]ai | AI inference and victim assessment |
| API endpoint | https://openrouter[.]ai/api/v1/chat/completions | LLM API requests |
| Service | ifconfig.me | Public IP discovery |
| Service | api.ipify.org | Public IP discovery |
| Service | rolimons[.]com | Roblox inventory valuation |
IOC handling note: The listed services are legitimate third-party services and should not automatically be classified as malicious. Detection should focus on suspicious processes, unexpected access patterns, the malware hash, and the context of the connections.
Cisco Talos did not identify dedicated attacker-owned C2 infrastructure for the observed VOLTA sample.
11. Detection and Threat Hunting
Security teams should focus on identifying credential-stealing behavior, suspicious browser execution, unauthorized access to application data, unusual API communication, and the creation of VOLTA-related forensic artifacts.
11.1 Endpoint Threat Hunting
Review EDR telemetry for:
Execution of suspicious or unknown Windows executables from user-writable directories.
Processes launching Chrome with the
--headlessargument.Unusual applications attempting to access browser credential stores.
Processes accessing Firefox, Thunderbird, Outlook, Slack, Discord, and Telegram data.
Access to cryptocurrency wallet directories and wallet vault files.
Execution of
WlanGetProfileListby unknown processes.Access to SSH and cloud configuration directories.
Creation of
Identity_Report.txtandAI_Analysis.txt.Creation of
volta_log.zip,volta_dox.zip, orvolta_doxnet.zip.Creation or modification of
%APPDATA%\volta_debug.txt.Unexpected access to Gmail, DoorDash, Uber Eats, Discord, and Roblox through authenticated sessions.
11.2 Network Threat Hunting
Review proxy, DNS, firewall, and network telemetry for:
Unexpected requests to
openrouter[.]aioriginating from non-approved applications.Unusual API requests to
/api/v1/chat/completions.Processes communicating with external services for AI inference without an established business requirement.
Unexpected access to
ifconfig.meandapi.ipify.org.Suspicious access to
rolimons[.]comfrom unknown processes.Large outbound transfers following browser credential or wallet data collection.
Multiple outbound connections to third-party services from an unfamiliar executable.
Because VOLTA uses legitimate third-party services, network indicators should be correlated with endpoint process information and user activity.
11.3 Microsoft Defender XDR – Suspicious Headless Chrome Execution
Cisco Talos recommends monitoring Chrome headless execution as part of detecting the App-Bound Encryption bypass technique.
The following query identifies Chrome processes launched with the --headless argument:
DeviceProcessEvents
| where Timestamp > ago(7d)
| where FileName in~ ("chrome.exe", "chrome")
| where ProcessCommandLine contains "--headless"
| project
Timestamp,
DeviceName,
AccountName,
FileName,
ProcessCommandLine,
InitiatingProcessFileName,
InitiatingProcessCommandLine,
SHA256
| order by Timestamp descInvestigate instances where Chrome is launched in headless mode by an unfamiliar process or from an unexpected execution path.
11.4 Microsoft Defender XDR – VOLTA Artifact Hunting
DeviceFileEvents
| where Timestamp > ago(30d)
| where FileName in~ (
"volta_debug.txt",
"Identity_Report.txt",
"AI_Analysis.txt",
"volta_log.zip",
"volta_dox.zip",
"volta_doxnet.zip"
)
or FolderPath contains @"\AppData\Roaming\"
| where FileName contains "volta"
or FileName contains "Identity_Report"
or FileName contains "AI_Analysis"
| project
Timestamp,
DeviceName,
ActionType,
FileName,
FolderPath,
SHA256,
InitiatingProcessFileName,
InitiatingProcessCommandLine
| order by Timestamp descThis query is intended to identify known VOLTA-related artifacts. The broader AppData path condition should be reviewed carefully and may be narrowed according to the environment's telemetry volume.
11.5 Microsoft Defender XDR – Known Sample Hash Hunting
DeviceProcessEvents
| where Timestamp > ago(30d)
| where SHA256 =~
"57ef4242666d78c50df145b34023df564f0fad38812d6a580e66433d427e2279"
| project
Timestamp,
DeviceName,
AccountName,
FileName,
FolderPath,
SHA256,
ProcessCommandLine,
InitiatingProcessFileName
| order by Timestamp descThe hash query can also be adapted for DeviceFileEvents to identify file creation or modification events involving the known sample.
Hunting note: These queries are detection starting points based on the published research. Validate the available table fields and telemetry coverage in the target Microsoft Defender environment before deploying them as production analytics rules.
12. Mitigation and Recommendations
12.1 Endpoint Protection
Deploy and maintain endpoint detection and response across Windows endpoints.
Monitor unknown applications launching Chrome with the
--headlessargument.Implement application control policies to restrict unauthorized processes from launching Chrome in headless mode.
Alert on suspicious access to browser credential stores, wallet files, and application session data.
Monitor unusual access to Windows WLAN profiles and stored wireless credentials.
Ensure endpoint protection is configured to detect and block information stealers.
12.2 Credential and Session Protection
Revoke potentially compromised browser and application sessions.
Reset passwords associated with affected accounts from a known-clean device.
Revoke Microsoft OAuth tokens and other exposed authentication tokens.
Rotate SSH keys, cloud credentials, API keys, and developer secrets that may have been accessed.
Review Gmail, Discord, Slack, and other application authentication activity for suspicious access.
Enable phishing-resistant MFA wherever supported.
12.3 Cryptocurrency Wallet Protection
Treat cryptocurrency wallet seed phrases stored in plaintext on a compromised endpoint as exposed.
Generate a new recovery seed on a trusted, air-gapped device.
Migrate funds to a newly created wallet.
Never reuse a potentially exposed recovery phrase.
Review transaction history for unauthorized transfers.
12.4 Browser and Application Security
Avoid storing sensitive credentials directly in browser password managers where stronger credential-management controls are available.
Use hardware-backed credential storage where supported.
Restrict access to browser profile directories.
Monitor unexpected processes attempting to access browser databases and session files.
Review suspicious OAuth application grants and revoke unauthorized access.
12.5 Incident Response
If VOLTA is suspected or confirmed:
Isolate the affected endpoint from the network.
Preserve
%APPDATA%\volta_debug.txtbefore remediation, if available.Collect EDR process telemetry, file artifacts, network logs, and authentication records.
Identify the scope of stolen credentials, sessions, wallet information, and cloud access.
Revoke exposed sessions and rotate affected credentials.
Investigate potential unauthorized access to cloud services and corporate applications.
Review cryptocurrency wallet activity and migrate exposed assets.
Rebuild the endpoint from a trusted image if system integrity cannot be established.
13. MITRE ATT&CK Mapping
The following techniques represent behaviors described in the Cisco Talos analysis. Some mappings should be treated as behavioral associations rather than proof that every technique was observed in every infection.
| Tactic | Technique | ID | Relevance |
|---|---|---|---|
| Credential Access | Credentials from Web Browsers | T1555.003 | Collection of browser-stored credentials |
| Credential Access | Steal Web Session Cookie | T1539 | Theft of browser and application session information |
| Credential Access | Steal Application Access Token | T1528 | Collection of OAuth and application authentication tokens |
| Credential Access | Credentials from Password Stores | T1555 | Collection of stored credentials and wallet-related secrets |
| Collection | Screen Capture | T1113 | Collection of screenshots |
| Discovery | System Information Discovery | T1082 | Collection of host and system information |
| Discovery | System Network Configuration Discovery | T1016 | Collection of network-related information |
| Collection | Data from Local System | T1005 | Collection of local files, profiles, and sensitive information |
| Collection | Email Collection | T1114 | Searching Gmail for receipts and personal information |
| Collection | Data from Cloud Storage | T1530 | Access to information through authenticated online services, where applicable |
| Exfiltration | Exfiltration Over Web Service | T1567 | Use of third-party web services for data processing and transfer |
14. Conclusion
VOLTA-AUTODOXER demonstrates an emerging evolution in information-stealing malware, where AI is integrated into the post-compromise workflow to automate the analysis and prioritization of stolen information.
By combining credential theft, session hijacking, cryptocurrency wallet collection, personal-data enrichment, and LLM-based victim assessment, the malware turns raw stolen data into structured intelligence that can help operators identify high-value targets.
The use of a multi-model fallback chain also indicates that the malware was designed to continue its assessment process when an AI model refuses a request.
For defenders, this incident reinforces the importance of monitoring not only traditional malware indicators but also suspicious access to browser profiles, application sessions, cloud credentials, and AI API services.
Organizations should prioritize endpoint detection, session revocation, credential rotation, browser security, and retrospective threat hunting to identify potential VOLTA infections and reduce the impact of information theft.
