2026-10-31 09:32 UTC
----------------
🎯 AI
===================
Cisco Talos has documented NIK C2, a new command-and-control framework that embeds a generative AI client in the malware itself. The implant is Go-based, runs on 64-bit Windows, and carries a full OpenAI API client exposed through an ai_generate command. Operators can send arbitrary prompts to the OpenAI Chat Completions API directly from the compromised host; the implant resolves API keys, assembles requests, and parses responses locally, then returns encrypted results to the actor's C2.
Architecture
Networking runs on a dual-channel design: TLS-secured WebSockets on port 443 as the primary channel, with an HTTP fallback polling loop firing every 250 milliseconds. All traffic is encrypted with AES-256-GCM derived from a hardcoded seed. For takedown resilience, traffic routes across three prioritized C2 tiers: runtime-updated URLs, an embedded encrypted server list, and a hardcoded fallback address.
The capability list is broader than data theft alone: credential dumping, keystroke streaming, RDP session hijacking, and process hollowing are all reported. There is also a C2-driven screen-locking ransomware module, with ransom demanded in Bitcoin or USDT.
The AI integration, realistically
The report describes capability, not a documented intrusion: the published text does not include a case where the model steered decisions in a specific incident. What is concrete is the plumbing. The API client is embedded, key resolution happens on the implant, and results return encrypted to the C2 rather than through the operator's browser. Talos argues that this collapses collection and exploitation into a single continuous session, erasing the offline processing delays that historically gave defenders detection windows. Given that NIK C2 also streams keystrokes and dumps credentials in the same session, the argument is at least coherent.
Versioning and OPSEC artifacts
Observed samples embed a build path, /opt/nik_panel/data/build_52b79724a1fc4b8f/main.go, which identifies the project as "nik_panel" and shows the implant was cross-compiled for Windows from a Linux environment. An embedded eight-byte hex token matches that build identifier, which Talos reads as the actor's infrastructure actively tracking and versioning implants.
All binaries were compiled with Go 1.23.4, released in December 2024, so the entire sample set dates no earlier than late 2024. PE TimeDateStamp fields are zeroed, and filesystem timestamps sit near epoch, manifesting as Nov 29, 1979. Talos built a timeline from first OSINT sightings of three distinct versions and infers, from the pace of capability addition, a technically proficient team with prior malware operations experience, actively expanding NIK C2 toward a full-service adversarial framework.
Infrastructure
One active server was observed in the wild at 87.120.196[.]120. Treat this as preliminary; the report is being published in stages, and no detection rules or hashes appear in the provided text.
Practical notes
Not from Talos, and not validated against samples: a quick hunt could correlate stable WebSocket sessions on port 443 with periodic HTTP calls at a 250-millisecond cadence from the same host, plus outbound calls to OpenAI API endpoints. Also worth watching for the screen-lock ransomware stage, since that converts a silent implant into a loud incident.
Talos positions continuous behavioral vigilance beyond traditional detection as the baseline. The concrete takeaway stands regardless of phrasing: generative API usage is now appearing inside malware workflows, and that surface is likely to keep growing.
🔹 NIKC2 #OpenAI #ThreatIntel #MalwareAnalysis #C2
🔗 Source: https://blog.talosintelligence.com/nik-c2-new-ai-enabled-framework/
Replies (0)
No replies.