Resources

Reports and research on AI security

In-depth reports from the Patronus team: what attacks on AI systems look like in practice, what the numbers say, and how organizations keep control. Free to download, no registration.

Two series, two jobs

Patronus Insights

The deep-dive series. One security topic per issue, worked through with scenarios, real incidents and concrete implementation, from OWASP risk coverage to AI governance in the workplace.

Patronus Radar

The AI security briefing for the DACH region. Orders the most important incidents, research and regulation of the past months, written for IT leads and decision-makers.

Patronus Insights · Episode 3

The road to sub-cent AI security inference

Sep 2026 · 12 pages · PDF · German · 1,0 MB

AI security becomes an inference problem as agents, tool calls, and context windows grow. This issue shows how Patronus Ark separates high-volume CPU routing from selective deep inference: at the validated operating point, roughly 25% of chunks reach L3 while F1 and false-positive behavior remain approximately at the level of full L3 inference. It connects that architecture to measured CPU and NVIDIA L4 throughput, TensorRT latency, and the planned Patronus Pro price of about €0.0067 per million protected tokens at continuous 4,000-token-per-second usage.

What's inside

Selective inference: why roughly 25% L3 promotion can retain approximately full-L3 quality at the validated operating point
Measured stack: about 963k CPU L2 input tokens/s and about 387k L3 tokens/s on a full NVIDIA L4
Economics and scaling: separate CPU and GPU bottlenecks, plus the assumptions behind about €0.0067 per million protected tokens

Download report (free)

Patronus Insights · Episode 2

OWASP for AI Security: which risks Patronus controls

Jun 2026 · 12 pages · PDF · German · 1,2 MB

OWASP maintains two separate top-10 lists for AI: the LLM Top 10 (2025) covering language-model inputs and outputs, and the Agentic ASI Top 10 (late 2025) covering autonomous agents that call tools, execute code, and connect to external systems via MCP. This report explains both lists, maps all 20 risk categories, and goes deep on five risks that Patronus Protect fully controls today, each with an explanation, a concrete attack example, and the technical enforcement via the action levels Allow, Flag, Redact, and Block.

What's inside

Why the LLM Top 10 and the Agentic ASI Top 10 address different layers, and must not be confused
Five risks in depth: Prompt Injection, Tool Misuse & Code Execution, identity and data abuse, Agentic Supply Chain, System Prompt Leakage
Coverage matrix: which of the 20 categories Patronus controls today, and what ships next

Download report (free)

Related on this site: AI Agent Security · AI Models · Blog: The MCP attack surface

Patronus Radar · Episode 1

AI Security Briefing: three months of AI security for DACH

Jun 2026 · 7 pages · PDF · German · 0,8 MB

Q2 2026 produced a string of concrete AI security incidents, from supply-chain attacks via Hugging Face and the LiteLLM library to zero-click vectors in AI browsers. This briefing groups the individual cases into three structural risks (compromised AI tooling, agent hijacking, shadow AI) and explains what the Digital Omnibus and current GDPR fines mean for companies in the DACH region.

What's inside

MCP under fire: architectural weaknesses plus a LiteLLM supply-chain incident affecting thousands of companies
Agent hijacking: when the productivity tool itself becomes the attacker
Shadow AI in DACH: visibility as the prerequisite for any control, plus Digital Omnibus and GDPR fines, classified

Download report (free)

Related on this site: Shadow AI Discovery · EU AI Act Assessment · Blog: Shadow AI blind spots

Patronus Insights · Episode 1

Managing AI in the workplace, without banning it

Jun 2026 · 8 pages · PDF · German · 1,2 MB

AI use in German companies more than doubled within a year: Bitkom 2026 puts productive adoption at 41%, up from 17%. At the same time, 39.7% of all AI interactions contain sensitive company data, and more than half run through private accounts, invisible to IT (Claude: 58.2%, Perplexity: 60.9%). IBM measured $670,000 in extra breach costs where shadow AI was involved. Using three scenarios: Chrome as an AI platform, whitelisting instead of banning, visibility as the foundation, this report shows how control works without a blanket ban.

What's inside

The numbers: Bitkom 2026, Cyberhaven's analysis of billions of real data flows, IBM Cost of a Data Breach 2025
Scenario Chrome AI mode: what happens when the default browser becomes an AI platform
Whitelist instead of ban: one approved AI, everything else visible and controlled

Download report (free)

Get new issues by email

New Insights and Radar issues land on the waitlist first, product updates included, unsubscribe any time with one click.

Join the release waitlist