Skip to main content

According to research by Cisco Talos published on 31 October 2026, VOLTA (also referred to as VOLTA-AUTODOXER) is described as an infostealer that incorporates a Large Language Model (LLM) to analyse exfiltrated data, reconstruct the victim’s profile and assess their potential for monetisation.

The distinctive feature is not the data collection itself, but the stage following exfiltration: credentials, sessions, personal information and traces of online activity are aggregated into a structured dossier and fed into an LLM, which assesses the victim’s financial potential and weaknesses in their operational security (OPSEC) practices.

AI is not used to generate malware or to automate the initial phase of the attack, but for the triage of targets after devices have been compromised.

 

LLM-based victim triage

Traditional infostealers collect data from compromised endpoints, organise it into archives and transmit it to operators, who can use it directly or resell it on the criminal markets. However, not all data is of equal value: a archive may contain hundreds of unusable passwords, expired sessions or personal information of little interest, alongside corporate credentials, still-valid authentication tokens, access to cloud infrastructure or cryptocurrency wallets. Manual selection requires time and expertise: VOLTA automates this task.

The malware sends the victim’s profile to an LLM service via OpenRouter’s APIs, with a prompt instructing the model to act as an analyst. The expected output comprises two sections: ‘WHAT THEY’RE WORTH’, classifying assets by their monetisation potential, and ‘THEIR OPSEC’, analysing the user’s security habits.

This is not merely a summary of the stolen files: the aim is to produce an assessment that can be used to determine which accounts, identities or pieces of information warrant further malicious activity. However, an assessment generated by an LLM does not guarantee that the economic value attributed to the victim is correct: the model may make inaccurate inferences or overestimate the relevance of certain elements. The main advantage for the attacker lies in the ability to carry out this selection automatically, without having to analyse each file individually.

 

Collected data from VOLTA

The data collection component features capabilities already familiar within the infostealer ecosystem, but covers a particularly wide range of applications and information. It targets:

  • passwords, session cookies and payment card details in Chromium-based browsers; data from Firefox, Thunderbird and Outlook profiles;
  • sessions and credentials from Discord, Telegram, Slack, Steam, Battle.net and Roblox;
  • Microsoft OAuth tokens, SSH keys, cloud and VPN configurations (gcloud, OpenVPN);
  • Wi-Fi passwords, screenshots and cryptocurrency wallet data, with a capability present in the analysed sample to retrieve login credentials for protected wallets, rather than simply copying the files.

 

Enrichment via compromised sessions

VOLTA can reuse compromised sessions to query online services and obtain additional information. Documented examples include searching for receipts in Gmail accounts, viewing order histories on delivery platforms and enriching Discord profiles.

These activities make it possible to link data which, when considered in isolation, might seem of little relevance: a receipt can help identify a physical address, a business account can reveal the employer, and document metadata can provide further details on a person’s identity. The result is a profile that may include geographical location, home address, organisation affiliation, partial financial information and references to the accounts used.

It is this correlation between digital identity and personal information that sets VOLTA apart from a simple credential harvesting tool.

 

LLM integration and fallback mechanism

VOLTA utilises the APIs of OpenRouter, a platform that provides access to various language models via a common interface.

In the sample analysed, the model configured as the first choice is anthropic/claude-opus-4.6. If the response contains expressions interpreted as a refusal, the malware can repeat the request using moonshotai/kimi-k2 and subsequently google/gemini-2.5-flash. The mechanism is relatively simple: the programme checks the text returned by the model for refusal phrases and, when it detects them, moves on to the next model.

This implementation shows that the malware author anticipated the possibility of requests being rejected by providers’ security systems. It does not, however, prove that the models produced usable assessments during genuine attacks, nor that the fallback mechanism is always effective.

The malware sends a structured set of information about the victim and requests a monetisation-oriented analysis. The response is saved to a file named AI_Analysis.txt, intended to be included amongst the collected results. The AI integration is also optional: in the absence of a configured API key, VOLTA can continue to function as a standard infostealer. The LLM component does not replace traditional information-stealing techniques, but adds an analytical capability to the process.

 

Indications of Infostealer-as-a-Service

The sample features several parameters that can be configured at compile time, including API keys and information relating to the Telegram channel used for exfiltration. This architecture suggests the possible existence of a builder capable of generating customised versions of the malware for different operators. If this business model were confirmed, the ability to produce automated dossiers could become accessible even to attackers with limited technical expertise.

However, it is necessary to distinguish between the characteristics of the code and conclusions regarding its use. The presence of configurable parameters does not, in itself, constitute proof of the existence of an actively marketed Infostealer-as-a-Service. Furthermore, the public information examined does not confirm the number of victims, the geographical spread of the infections or the initial vector used to distribute VOLTA.

The case therefore represents a concrete example of the integration of malware and AI, not evidence of a criminal campaign already operating on a large scale.

 

Risk to corporate identities

For organisations, the most significant risk is not necessarily the theft of the password itself, but the compromise of the digital identity as a whole.

A personal device also used for work may contain SaaS sessions, cloud credentials, OAuth tokens and information relating to the organisation. If these elements are stolen, an attacker could attempt to reuse sessions that are still valid or exploit the credentials to gain access to corporate resources. Under certain conditions, the theft of a session token may allow access to a service without having to repeat the entire multi-factor authentication (MFA) process: whether this is actually possible depends on the safeguards implemented by the service, the token’s validity and session controls.

VOLTA adds to this scenario the automated correlation between personal information and professional access: knowing that a person works for a particular company, uses specific collaborative platforms and holds credentials associated with cloud services can make them a more attractive target for subsequent activities. The malware could therefore help facilitate the selection of potential targets for account takeover, Business Email Compromise (BEC), spear phishing campaigns or attempts to gain initial access to corporate networks. These are risk scenarios consistent with documented capabilities, not consequences already attributed to an observed VOLTA campaign.

 

Detection and Incident Response

Network Telemetry

As the analysis request is generated by the compromised device, traffic to OpenRouter may constitute an anomalous signal, particularly when it originates from endpoints that should not be directly utilising Artificial Intelligence APIs. However, the destination alone is not sufficient to identify an infection: it is necessary to correlate the connections with the process generating them, the endpoint’s context and the immediately preceding activities.

 

Endpoint Telemetry

The analysis describes a data collection mechanism that utilises launching Chrome with the –headless parameter to bypass App-Bound Encryption protections under certain conditions. Headless execution is not malicious in itself, as it is also used by legitimate automation tools, but it becomes of interest when initiated by unknown processes or from directories not normally associated with authorised software.

 

Forensic artefacts and IOCs

Artefacts described: %APPDATA%\volta_debug.txt (retains information on operations even after the executable has self-deleted), Identity_Report.txt, AI_Analysis.txt, volta_log.zip, volta_dox.zip and volta_doxnet.zip. SHA-256 hash of the documented sample:

57ef4242666d78c50df145b34023df564f0fad38812d6a580e66433d427e2279

These indicators can support targeted checks, but a detection strategy based solely on hashes or filenames risks failing to identify any variants. For this reason, it is preferable to combine EDR telemetry, process creation, anomalous access to browser data, connections to external APIs and suspicious activity on identity services.

 

Response and remediation

In the event of a confirmed compromise, the response must consider not only accounts as potentially compromised, but also active sessions, API keys, cloud credentials and other secrets present on the device. Simply changing the password may not be sufficient. Where supported, compromised sessions must be revoked, tokens invalidated, keys rotated and any access attempts following the theft verified.

 

In conclusion

VOLTA does not introduce a completely new technique for stealing information: most of its data collection capabilities belong to a category of malware that is already widely known. What is new is the role assigned to Artificial Intelligence: the language model becomes a component of the criminal process tasked with interpreting information, setting priorities and reducing the operator’s workload (use of an LLM).

Looking ahead, this approach could facilitate greater automation in victim selection, making the exploitation of large quantities of stolen data more efficient. This does not necessarily mean that attacks will become more sophisticated at every stage, but rather that part of the analysis which previously required human intervention can be incorporated directly into the malware.

For organisations, protecting credentials remains essential, but it is not enough: it is necessary to strengthen session security, identity management, endpoint control and the ability to detect anomalous behaviour before the stolen information is used for new attacks.

The VOLTA case illustrates how AI can transform a repository of stolen data into a reasoned list of potential targets. It is this ability to link the theft of information to its potential monetisation that makes it a significant case for the evolution of cybercrime.

Analysis by Vasily Kononov – Threat Intelligence Lead, Cyberoo

 

References

  1. Cisco Talos – Ryan Fetterman, Appraisal by AI: Model-driven victim prioritisation in the VOLTA infostealer, 31 October 2026.
    Appraisal by AI: Model-driven victim prioritisation in the VOLTA infostealer
  2. Cisco Talos, The VOLTA infostealer: Outsourcing cybercrime grunt work to AI, 4 November 2026.
    The VOLTA infostealer: Outsourcing cybercrime grunt work to AI – Talos Takes
  3. Cloud Security Alliance, Appraisal by AI: VOLTA and LLM-Driven Infostealer Victim Triage, 2 November 2026.
    Appraisal by AI: VOLTA and LLM-Driven Infostealer Victim Triage – Lab Space
  4. WIRED, Spooky New Infostealer Malware Uses AI to “Autodox” Victims, November 2026.
    Spooky New Infostealer Malware Uses AI to ‘Autodox’ Victims