MAL-2026-20196
Malicious code in character-ai-sdk (npm)
Auswirkung
character-ai-sdk@1.0.0 presents itself as a persona/memory SDK but its main entry (index.mjs → personas.mjs) top-level imports./cache-warm.mjs, a ~300 KB obfuscator.io-style blob (hex-renamed string array of base64+XOR entries) with no library-shaped exports. Decoded property-access artifacts in the blob reference `process.exit`, `process.pid`, hostname, and child-process APIs, consistent with runtime-resolved process control and host enumeration executed on `require`/`import`. Three additional obfuscated siblings of the same shape (build-cache.mjs ~200 KB, cache-sweep.mjs ~290 KB, workspaces.js ~280 KB) are staged next to the main-reachable loader. The tarball also ships platform-specific native addons at native/{linux,darwin,win32}/{x64,arm64}/cache-probe.node whose exported symbols are a vendored VM-detection library (VMDetect, VM::qemu_fw_cfg, VM::vbox_module, VM::hypervisor_bit, VM::smbios_vm_bit, VM::dmidecode, VM::chassis_vendor, VM::mac_address_check), confirmed by native/LICENSE-node-vm-detect.txt. A persona-lookup SDK has no legitimate need for hypervisor/QEMU/VirtualBox/DMI/MAC-based sandbox detection, and the combination — heavy obfuscated loader reachable from the main entry, multiple identical-shape dropper siblings staged alongside it, and cross-platform anti-sandbox native addons — matches the shape of a staged, sandbox-evading dropper that executes on module load.
Betroffene Pakete
| Ökosystem | Paket | Betroffen | Behoben in |
|---|---|---|---|
| npm | character-ai-sdk | unbekannt | nicht erfasst |