Security researchers recently identified a piece of malware named PromptLock that uses a local artificial intelligence model to help create and run harmful code on infected machines. The finding comes from ESET researchers and has been reported by multiple security outlets; investigators say PromptLock can scan files, copy or steal selected data, and encrypt user files, with code for destructive deletion present but not active in analysed samples.
What does “AI-powered” mean here?
Instead of a human writing every malicious script in advance, PromptLock stores fixed text prompts on the victim machine and feeds them to a locally running language model. That model then generates small programs, written in the lightweight Lua language, which the malware executes immediately. Researchers report the tool uses a locally accessible open-weight model called gpt-oss:20b through the Ollama API to produce those scripts. Because the AI runs on the infected computer rather than contacting a remote service, the activity can be harder to spot.
How the malware works
According to the technical analysis, PromptLock is written in Go, produces cross-platform Lua scripts that work on Windows, macOS and Linux, and uses a SPECK 128-bit encryption routine to lock files in flagged samples. The malware’s prompts include a Bitcoin address that investigators linked to an address associated with the pseudonymous Bitcoin creator known as Satoshi Nakamoto. Early variants have been uploaded to public analysis sites, and ESET treats this discovery as a proof of concept rather than evidence of widespread live attacks.
Why this matters
Two features make this approach worrying for defenders. First, generated scripts vary each time, which reduces the effectiveness of signature or behaviour rules that rely on consistent patterns. Second, a local model produces no network traces to cloud providers, so defenders lose one common source of detection and takedown. Together, these traits could make automated malware harder to detect and classify.
Practical, plain steps to protect yourself:
1. Do not run files or installers you do not trust.
2. Keep current, tested backups offline or on immutable storage.
3. Maintain up-to-date operating system and antivirus software.
4. Avoid running untrusted local AI models or services on critical machines, and restrict access to local model APIs.
These steps will reduce the risk from this specific technique and from ransomware in general.
Bottom line
PromptLock is a clear signal that attackers are experimenting with local AI to automate malicious tasks. At present it appears to be a work in progress and not an active campaign, but the researchers stress vigilance and standard defensive practices while security teams continue monitoring developments.
According to ESET researcher Lukas Stenfanko who examined a sample after getting a tip from MalwareHunterTeam, it was found that one of the noteworthy new features seen in the most recent GravityRAT version is the ability to collect WhatsApp backup files.
A remote access tool called GravityRAT has been used in targeted cyberattacks on India since at least 2015 and is known to be in use. There are versions for Windows, Android, and macOS, as previously reported by Cisco Talos, Kaspersky, and Cyble. However it is still unknown who is the actor behind GravityRAT, the group has been internally defined as SpaceCobra.
Although GravityRAT has been active since at least 2015, it only began specifically focusing on Android in 2020. Its operators, 'SpaceCobra,' only employ the malware in specific targeting tasks.
According to ESET, the app is delivered via “bingechat[.]net” and other domains or distribution channels, however, the downloads require invites, entering valid login information, or creating a new account.
While registrations are currently closed, this method only enables the threat actors to distribute the malware to targeted users. Additionally, accessing a copy for analysis becomes more difficult for researchers.
Upon installation on the target's smartphone, the BingeChat app makes dangerous requests for access to contacts, location, phone, SMS, storage, call records, camera, and microphone.
Since these are some typical permissions asked of the users for any instant messaging apps, the malicious app goes unsuspected.
The program provides call records, contact lists, SMS messages, device location, and basic device information to the threat actor's command and control (C2) server before the user registers on BingeChat.
Along with the aforementioned records, files, and document files of jpg, jpeg, log, png, PNG, JPG, JPEG, txt, pdf, xml, doc, xls, xlsx, ppt, pptx, docx, opus, crypt14, crypt12, crypt13, crypt18, and crypt32 types, have also been compromised.
While SpaceCobra’s malware campaign is mainly targeting India, all Android users are advised to refrain from downloading APKs anywhere other than Google Play and be very careful with potentially risky permission requests while installing any app.