Research & Intelligence
Four streams of published work from active engagements — sanitized findings, named vulnerabilities, no client data. We publish what we learn so the field gets safer, faster.
Indirect Prompt Injection 2026: A Field Survey
Six months of production telemetry across 40+ AI deployments. How automated scrapers are weaponizing RAG pipelines against internal LLMs, and the three defensive patterns that actually work.
Real-world AI vulnerability case studies from active engagements. Each report names an attack class, shows how it manifested in production, and documents the remediation that worked.
Indirect Prompt Injection 2026
Six months of field data on RAG weaponization.
Tool-Call Chain Privilege Escalation
How agent delegation undermines trust boundaries.
Embedding Poisoning at Scale
Document-layer injection in production retrieval pipelines.
Public log of AI logic leaks and vulnerabilities discovered and disclosed by LogicLeak researchers. Each entry documents affected system class, technical description, and remediation guidance — never client names without consent.
Embedding Poisoning via Crafted Markdown in Hosted Vector Indexes
Cross-Tenant Retrieval Leakage in Multi-Tenant RAG Deployments
Tool-Call Chain Bypass in Popular Agent Framework
Prompt-Layer Auth Bypass in AI Customer Support Platform
How we approach adversarial security for AI systems. Our methodology is open — we publish what we believe so clients and researchers can challenge it, and so the field can build on it.
The Adversarial Probing Methodology
Our 12-stage process for red-teaming production AI systems, from reconnaissance to remediation handoff.
Severity Scoring for AI Vulnerabilities
How we rank AI security findings against CVSS, OWASP LLM Top 10, and engagement-specific risk models.
Rules of Engagement
Our operational ethics, compliance, and testing boundaries — what we will and won't do on engagements.
Sanitization & Disclosure Policy
How we strip client identifiers from published research while preserving technical fidelity.
A continuously updated map of the attack patterns we see emerging across multi-agent, RAG, and tool-calling AI systems. Less a paper, more a worldview.
2026
The 2026 Agentic Threat Landscape
Across Q1–Q2 2026, we've seen the attack surface shift decisively toward multi-agent privilege escalation, indirect injection via tool outputs, and supply-chain attacks on shared agent frameworks. This living document tracks what we're seeing, what we expect next, and how defenders should prioritize.
- 1. Multi-agent privilege escalation
- 2. Indirect injection via tool outputs
- 3. Agent framework supply chain
- 4. RAG perimeter degradation
Get research before it goes public
One email per month. New threat reports, vulnerability disclosures the moment they're published, and pre-release access to upcoming papers. Unsubscribe anytime.