Skip to main content
Quarterly research

ShieldPi Research

Anthropic publishes safety research and owns that conversation. We do the same for offensive AI security. Quarterly papers, real findings, fully reproducible methodology — submitted to USENIX, BlackHat, DEFCON, and IEEE S&P.

Published & Draft

2026 Q2Draft·BlackHat Webcast Q3 2026 / Industry briefing· 12 min read

Post-Quantum AI Risk: Why Your LLM Security Strategy Needs a Crypto-Agnostic Layer

Two unrelated forcing functions are about to collide in enterprise security budgets — generative AI agents in production and quantum-relevant cryptographic breaks. This paper argues for a crypto-agnostic AI security layer.

ShieldPi ResearchRead paper

Upcoming

2026 Q3Outline·USENIX Security 2026

Multi-Turn Authority Escalation against GPT-5 Enterprise

A formal study of the 5-layer authority escalation pattern (curious learner → researcher → official → regulator → emergency) against frontier enterprise models. Includes ASR breakdown per layer and per attack category.

Outline — full draft TBA

2026 Q4Outline·DEFCON AI Village 2026

Knowledge Graph–Driven Adaptive Red-Teaming (V6 architecture)

How we built a living knowledge graph of attack-to-model edges that learns within a single scan and across the customer base. Includes the mid-scan learning loop architecture.

Outline — full draft TBA

Publication cadence

Quarterly

One technical paper to a top-tier venue (USENIX, BlackHat, DEFCON, IEEE S&P) covering a real finding from the corpus.

Monthly

Short technical posts on the blog summarizing notable jailbreaks added to the harvester.

Continuous

The Living Leaderboard at shieldpi.info updates weekly with delta from the harvester + customer scans.

All findings reproducible from our public methodology endpoint.

Open methodology →Live intelligence →research@shieldpi.io