Red Specter Specter Social is an LLM-powered social engineering assessment platform designed to simulate adversarial human-targeting attacks across multiple communication channels. Unlike traditional phishing simulators that rely on static email templates, it uses an autonomous reasoning engine to build psychological profiles of targets and adapt attack strategies in real time. The platform operates across six channels: email (SMTP spear phishing), voice (TTS vishing with voice cloning), SMS (smishing), web (credential harvesting portals), chat (Teams/Slack impersonation), and video (deepfake avatars). Core architecture consists of four components: - Human Target Model: builds psychological profiles including authority mapping, stress triggers, and compliance patterns - Decision Engine: LLM-powered planning that selects attack type, channel, persona, and timing - Goal Decomposer: breaks objectives into incremental, low-suspicion sub-goals across multiple interactions - Resistance Engine: analyses target pushback and generates countermeasures, pivoting channel, persona, or timing The platform supports 10 attack types including spear phishing, vishing, pretexting, C-suite impersonation, IT support impersonation, vendor impersonation, urgency escalation, multi-channel persistence, goal decomposition, and consensus manufacturing. Engagements follow a 7-phase loop: Recon, Profile, Plan, Execute, Adapt, Escalate, and Report. OSINT collection feeds the profiling phase. Reports are Ed25519-signed and exportable to SIEM in JSON and PDF formats. Two operational modes exist: Standard (simulated deepfake/credential harvesting) and Unleashed (live voice cloning, live deepfake video, live credential harvesting), with the latter requiring an Ed25519 key gate. LLM backends supported: Ollama (local, air-gapped), OpenAI GPT-4o, and Anthropic Claude. Installed via pip.