RS-2026-093: TORPEDO v2.2.0 — Autonomous Destruction with Adaptive Fallback
Year: 2026
Type: preprint
Abstract: This paper presents the live fire validation of SPECTER TORPEDO v2.2.0 (T288), a WMD-class autonomous destructive agent developed under the NIGHTFALL offensive AI security research framework. Two separate engagements were conducted against an Ollama AI inference server deployed in an isolated Docker network with no external target connectivity, producing cryptographically signed evidence records across both runs.The first engagement (TPD-df7583ea3bb1, RS-2026-070) demonstrated full HTTP-layer destruction of 21/21 components across the target stack. The second engagement (BBX-9e395ed810f5) revealed a critical architectural characteristic: when TORPEDO's primary HTTP DELETE vector is rejected by the target (HTTP 405), the agent autonomously adapts to a container-level execution fallback, achieves verified model destruction, and continues the full kill chain including DEADMAN autonomous fire activation without operator intervention.Post-engagement verification via ollama list confirmed zero models present following a confirmed pre-engagement load of llama3.2:1b (1.3 GB). We are not aware of any detection capability for this adaptive destruction pattern in currently deployed AI infrastructure monitoring tooling. The absence of a defensive countermeasure represents a critical gap in enterprise AI security architecture.
Author: Barron, Richard
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225393382