InfoResearchPeer-reviewedLLM-specific
Telemetry Ablation for Post-Incident Prompt-Injection Reconstruction: Evidence Sufficiency and Human Source Attribution in an LLM–Integrated Application
- Published
- Record updated
Summary
Researchers model prompt-injection incidents in LLM-integrated applications as a typed provenance multigraph and test whether telemetry supports post-incident reconstruction. Seven investigators reconstructed 168 synthetic incidents across seven evidence arms, and a pre-frozen minimum forensic logging profile rule returned ABSTAIN. With complete telemetry, detection recall was 0.921, channel-level attribution 0.688, and exact object attribution 0.176, and no empirical minimum profile was identified.
Related items
- HighGHSA-8vf2-hxg5-crqh: Kubeshark: MCP `start_kubeshark` argument injection via `pod_regex` → image redirect → cluster takeoverSimilar attack · GitHub Advisory Database
- MediumCVE-2026-126681: AWS Labs Timestream for InfluxDB MCP Server exposes API tokenSimilar attack · NVD/CVE Database
- HighCVE-2026-126681 - Improper restriction of intended endpoints in AWS Labs Timestream for InfluxDB MCP ServerSimilar attack · AWS Security Bulletins
- InfoIn Other News: New CISA Director Nominee, Connected Car Hacking, Baker Tilly Data BreachSimilar attack · SecurityWeek
- MediumHow one AI agent exposed five critical enterprise security gapsSimilar attack · BleepingComputer