T3MP3ST is a self-hosted, multi-agent offensive security framework that wraps existing AI coding agents—Claude Code, Codex, Hermes, OpenCode, Oh My Pi, or local models via Ollama, LM Studio, and vLLM—into an autonomous kill chain for authorized testing. It operates keylessly by default, requiring no new API keys or cloud tenancy, and exposes a browser-based War Room (port 3333) and HTTP/MCP APIs. The framework implements an 8-operator cell (Recon, Scanner, Exploiter, Infiltrator, Exfiltrator, Ghost, Coordinator, Analyst) mapped to MITRE ATT&CK phases, though only the single-agent Recon path is currently benchmarked; coordinated swarm exploitation remains unproven. Headline results, all re-derivable via npm run verify-claims from committed artifacts, include 90.1% pass@1 (Wilson-95 CI 86.2–92.9) on XBOW’s 104-challenge XBEN suite (surpassing XBOW’s self-reported 85%), 23/40 (58%) hint-free solves on Cybench, and 8/10 exact file/line/CWE matches on 10 held-out post-cutoff CVEs (CVE-Zero) across seven languages. The Arsenal provides 36 built-in tools (111 with T3MP3ST_FULL_ARSENAL ), including nmap, Ghidra, and Metasploit adapters, with egress-scope containment enforced by default. A coordinated-disclosure pipeline (OSV novelty check, live PoC, refuter panel, CVSS) is stable and has produced vendor drafts. Domains such as cloud, mobile, binary, and AD are scaffolded but unbenchmarked. The project is AGPL-3.0 licensed and emphasizes reproducible measurement, honest scope labeling, and keyless operation.