AI Security Insights

The on-device AI security research you need.

Practical guides, threat research, and architecture insights for security teams building AI-safe organisations.

Featured

Detecting risky shell commands

Oct 7, 2026 · 16 min

Read Article

Latest Articles

PDF Prompt Injection: Scan Before RAG and Agents

A PDF can contain instructions aimed at an AI. Understand the role of extraction, RAG and the controls around document processing and agent actions.

Agent Security · Oct 7, 2026 · 5 min · Read →

Manus Prompt Injection: Warning After Code Execution

The Manus attack triggered a warning after code execution. What that means for input screening, tool approvals and limited permissions in your AI agents.

Agent Security · Oct 6, 2026 · 5 min · Read →

Lakera vs. Patronus: Comparing prompt injection detection

Lakera vs. Patronus and PromptGuard vs. Patronus on 825 shared inputs: Patronus produced fewer false positives, while Lakera detected more attacks in selected agent benchmark texts.

Threat Research · Sep 27, 2026 · 7 min · Read →

Wolf Defender on the Buried Injections benchmark

Wolf Defender Small v2 detects 524 of 629 embedded attacks with one false positive. Two internal candidates detect 628 and 629. What matters is what the benchmark actually tests.

Threat Research · Sep 25, 2026 · 4 min · Read →

Scan an MCP server before connecting your AI agent

An MCP server supplies the tool names, descriptions, and schemas an agent reads. Inspect that metadata before the server joins the agent's tool catalog.

Agent Security · Sep 24, 2026 · 6 min · Read →

Introducing the Patronus Scanner API: check prompts, documents and MCP servers with one call

The same detection that runs on the device in Patronus Protect is now an API: one endpoint for text, URLs, files and MCP servers, returning a verdict, evidence spans and confidence. 100 scan units a day without an account, then free with an API key.

Agent Security · Sep 19, 2026 · 7 min · Read →

Sub-cent AI security inference: Why only 25% needs the deep model path

At its validated operating point, Patronus Ark sends only about 25% of chunks to deep L3 inference while remaining approximately at full-L3 quality.

Architecture · Sep 19, 2026 · 4 min · Read →

Safety before speed: what Altman, Amodei and Musk are now calling for

Three of the best-known figures in AI agree on one thing: model capabilities must not outrun our ability to control them safely. For companies, this is a debate about securing agents.

Agent Security · Sep 14, 2026 · 6 min · Read →

How NTDBs Make Decisions Without a Transformer

How Patronus Ark uses shared embeddings and a fusion of inexpensive classifiers to make Layer-2 security decisions without running a transformer.

Architecture · Sep 12, 2026 · 10 min · Read →

AI DLP vs. classic DLP: why your DLP can't see prompts to local models

The most common sentence in conversations with security teams is: we have DLP, but it doesn't see the prompts. That is not a configuration error. It is an architecture question.

Architecture · Sep 7, 2026 · 7 min · Read →

AI security vendors in the DACH region 2026: who does what

Search for AI security in the DACH region and you get a list of platforms, governance tools, model providers and a handful of actual security products. This overview sorts them by what they really do.

Architecture · Sep 7, 2026 · 9 min · Read →

Langdock vs. Patronus Protect: AI workspace or AI firewall?

Both products come up in the same conversations but solve different problems. Langdock gives employees a safe place for AI. Patronus shows whether they use it, and protects every other path.

Architecture · Sep 7, 2026 · 8 min · Read →

What is prompt injection? Definition, examples and defenses

Prompt injection has topped the OWASP list for LLM applications for years. With agents that read web pages and execute tools, a chat trick has turned into an attack on company data.

Agent Security · Sep 7, 2026 · 9 min · Read →

What is shadow AI? Definition, risks, examples and detection

Every company knew shadow IT. Shadow AI is faster, quieter and lives inside tools that were approved long ago. A primer with definition, examples and a comparison of the discovery methods.

Shadow AI · Sep 7, 2026 · 8 min · Read →

Why your firewall, proxy and CASB miss AI tools running on laptops

The question comes up in almost every first call: we have a firewall, a proxy and a CASB, so why can't we see the AI usage? The answer lies in four blind spots all three controls share.

Shadow AI · Sep 7, 2026 · 6 min · Read →

Wolf Defender v2: prompt injection detection that runs on the laptop

Wolf Defender v2 detects prompt injection and jailbreaks in German and English, with a 2,048-token context, on a laptop CPU. The biggest change since v1: far fewer false positives on hard benign text.

Threat Research · Sep 7, 2026 · 8 min · Read →

How We Run Real-Time AI Security Inference On-Device on a CPU

Patronus Ark runs layered AI security inference on an ordinary laptop CPU at roughly 750 MB peak RSS and around 500 milliseconds median latency.

Architecture · Aug 31, 2026 · 14 min · Read →

CoSnitch (CVE-2026-24301): One Click, and Copilot Reads Your Gmail and Drive

A single crafted link was enough to drain data from connected Gmail, Drive and calendar accounts. With no visible confirmation. The flaw is patched. The underlying problem for companies is not.

Shadow AI · Aug 24, 2026 · 7 min · Read →

Securing AI Coding Agents: What GhostApproval, Ghostjacking and CVE-2026-62830 Have in Common

Three security incidents in six weeks, three different vendors, one shared pattern: there is a gap between what an AI agent actually does and what a human gets to see. That gap is where attackers operate.

Agent Security · Aug 24, 2026 · 9 min · Read →

Our AI-Security Model Zoo Is Now Open Source

Threat Research · Jul 21, 2026 · 7 min · Read →

Orca-Sonar: Our Multilingual Document Classifier for AI Security

We're releasing Orca-Sonar, a fast, multilingual model (German + English) that sorts text into 7 topic classes. It's part of the Patronus Protect stack and routes incoming content by topic and sensitivity before it reaches an LLM, a DLP gate, or a storage tier.

Threat Research · Jun 12, 2026 · 8 min · Read →

When Malware Starts to Think: Why AI-Driven Worms Require a New Security Layer

A new research paper shows how AI agents can turn traditional computer worms into adaptive, self-replicating attack systems. Instead of following a fixed exploit chain, these worms can reason, use tools, learn from failed attempts, and spread across networks by using compromised compute. This shift makes AI runtime security essential, and shows why Patronus Protect is built for the next generation of AI-driven malware.

Threat Research · Jun 9, 2026 · 6 min · Read →

EU AI Act Article 15: 16 Extra Months Just Landed. Here Is What to Do With Them.

Article 15 of the EU AI Act becomes binding on August 2, 2026. Penalties hit 15M EUR. Here is what the law requires and how to check your gaps.

EU AI Act · May 21, 2026 · 12 min · Read →

How Patronus Detects AI Traffic Across WebSockets, SSE, gRPC, Protobuf and Modern Encodings

How modern AI traffic increasingly moves beyond simple JSON APIs into WebSockets, SSE, gRPC, Protobuf and binary streaming protocols, and how Patronus transparently detects and analyzes AI interactions across these transport and encoding layers while intentionally separating AI detection from semantic extraction.

Architecture · May 17, 2026 · 7 min · Read →

Agentic AI Broke Your Security Model

Your IAM was built for humans. Your DLP was built for clicks. Your audit log was built for human-pace actions. Agentic AI breaks all three. In 2026, 88% of enterprises had AI-agent incidents. Here is what changed and what real agent-aware security has to do.

Agent Security · May 9, 2026 · 11 min · Read →

Your security stack can't see Shadow AI

Your CASB sees the SaaS. Your DLP sees encrypted bytes. Your firewall sees a destination. None of them see Shadow AI: prompts pasted into ChatGPT, coding agents calling internal Jira, browser extensions reading confidential docs. Here is why the stack is blind, and what fixes it.

Shadow AI · May 6, 2026 · 9 min · Read →

MCP the new attack surface

MCP turns LLM clients into orchestrators that read Gmail, write to Jira, execute shell commands. The average dev laptop now runs three to eight MCP servers, most installed without IT approval. No traditional security tool can see them. Here is the attack surface and what fixes it.

Agent Security · Apr 22, 2026 · 8 min · Read →

New AI-Security Model: Wolf Defender

Threat Research · Apr 15, 2026 · 3 min · Read →

Cloud vs. On-Device AI Security

Most AI security tools route every prompt through the vendor's cloud to inspect it. You added a tool to reduce data exposure and ended up with two cloud companies holding your prompts instead of one. Here is what that costs and what the alternative looks like.

Architecture · Apr 2, 2026 · 8 min · Read →

Today we’re releasing our first security model: Wolf Defender

Threat Research · Mar 18, 2026 · 5 min · Read →

Patronus Protect - On-Device Protection for your AI Interactions

Architecture · Jan 18, 2026 · 7 min · Read →

Try Patronus on your own computer.

Pick a case in the live demo and see straight away what reaches the provider.

Go to release Contact

Reports and newsletter →