<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>AI Agent Security Guides, Standards, and MCP Security on PipeLab</title><link>https://pipelab.org/learn/</link><description>Practical guides and open standards for AI-agent security, including MCP defense, evidence verification, compliance, and secure agent setup.</description><generator>Hugo</generator><language>en-US</language><lastBuildDate>Mon, 16 Nov 2026 22:42:28 +0000</lastBuildDate><atom:link href="https://pipelab.org/learn/index.xml" rel="self" type="application/rss+xml"/><item><title>MCP Security Best Practices: 2026 Checklist</title><link>https://pipelab.org/learn/mcp-security/</link><guid>https://pipelab.org/learn/mcp-security/</guid><pubDate>Tue, 24 Feb 2026 00:00:00 +0000</pubDate><description><![CDATA[<p><em>Last updated: 2026-11-14.</em></p>
<p>MCP security is the practice of protecting AI agents that use the Model Context Protocol from threats originating in tool descriptions, tool arguments, tool responses, and the underlying transport. The four channels each carry different risk: tool descriptions and responses can carry hidden instructions, tool arguments can leak credentials, and transport metadata can hide exfiltration. Static install-time scanners catch some risks but miss rug-pull attacks where a server changes its tools mid-session, so runtime inspection at the proxy layer is necessary.</p>]]></description></item><item><title>MCP Vulnerabilities: Known Risks and Defenses</title><link>https://pipelab.org/learn/mcp-vulnerabilities/</link><guid>https://pipelab.org/learn/mcp-vulnerabilities/</guid><pubDate>Tue, 07 Apr 2026 00:00:00 +0000</pubDate><description><![CDATA[<p>MCP vulnerabilities are real and growing. The Model Context Protocol gives AI agents access to external tools, and those tools introduce attack surface that did not exist when agents only processed text. Every MCP connection is a trust boundary, and most agents cross it without any inspection.</p>
<p>This page is the canonical reference for known MCP vulnerabilities. It catalogs the attack classes, lists the public CVEs, cites primary sources, and names the defenses that apply to each class. For the full ecosystem snapshot, see the <a href="/blog/state-of-mcp-security-2026/">State of MCP Security 2026</a> report.</p>]]></description></item><item><title>MCP Proxy: How It Scans MCP Traffic Bidirectionally</title><link>https://pipelab.org/learn/mcp-proxy/</link><guid>https://pipelab.org/learn/mcp-proxy/</guid><pubDate>Tue, 07 Apr 2026 00:00:00 +0000</pubDate><description><![CDATA[<p>An MCP proxy sits between your AI agent and its MCP servers. It intercepts every JSON-RPC message flowing in both directions, scanning tool descriptions, tool call arguments, and tool responses before they reach the agent or the server. The agent connects to the proxy. The proxy connects to the real MCP server. Everything flows through the scanning pipeline.</p>
<p>If you need the category definition first, read <a href="/learn/what-is-mcp/">What Is MCP?</a>. This page assumes you already know the protocol and want the runtime control layer.</p>]]></description></item><item><title>AI Agent Security &amp; Hardening: 2026 Checklist Mapped to OWASP</title><link>https://pipelab.org/learn/ai-agent-security/</link><guid>https://pipelab.org/learn/ai-agent-security/</guid><pubDate>Tue, 07 Apr 2026 00:00:00 +0000</pubDate><description><![CDATA[<p><em>Last updated: 2026-11-14.</em></p>
<p>AI agent security is not one tool or one layer. Your agent has API keys, file system access, and an internet connection. When something goes wrong, secrets leave through HTTP requests, MCP tool calls, or DNS queries before anyone notices. The model&rsquo;s built-in safety training helps, but it&rsquo;s not a security boundary. It can be bypassed with the right prompt.</p>
<p>Real defense requires multiple layers, each catching different attacks at different points. Most teams either have zero layers or think one tool covers everything. Neither is true.</p>]]></description></item><item><title>Pipelock Action Receipt Format (Implementation Spec)</title><link>https://pipelab.org/learn/action-receipt-spec/</link><guid>https://pipelab.org/learn/action-receipt-spec/</guid><pubDate>Thu, 09 Apr 2026 00:00:00 +0000</pubDate><description>&lt;p>&lt;a href="/pipelock/">Pipelock&lt;/a> emits mediator-signed action receipts from outside the
agent trust boundary. The action receipt is the primitive that backs that
claim: a self-contained, Ed25519-signed record that the Pipelock mediator recorded
an attempted action and adjudicated it, signed from outside the agent trust
boundary. Each receipt records what happened, who authorized it, which policy
governed it, and what decision the mediator made. Receipts are chained by
hash, so tampering with any one of them breaks every receipt that follows.&lt;/p></description></item><item><title>Best MCP Security Tools 2026: Scanners, Gateways, Proxies Compared</title><link>https://pipelab.org/learn/mcp-security-tools/</link><guid>https://pipelab.org/learn/mcp-security-tools/</guid><pubDate>Sat, 11 Apr 2026 00:00:00 +0000</pubDate><description><![CDATA[<p>MCP (Model Context Protocol) gives AI agents access to external tools. Each connection is a trust boundary. For the full <a href="/learn/mcp-security/">MCP security</a> threat model and best-practices baseline, start with the pillar guide. A compromised or malicious MCP server can poison tool descriptions, steal credentials through tool arguments, or inject prompts in responses.</p>
<p>The ecosystem has responded with MCP security tools. They vary widely in what they actually check. This guide breaks down the categories, names the major tools, and helps you evaluate what you need.</p>]]></description></item><item><title>MCP Gateway Open Source vs Commercial: Comparison and How to Choose</title><link>https://pipelab.org/learn/mcp-gateway/</link><guid>https://pipelab.org/learn/mcp-gateway/</guid><pubDate>Sat, 11 Apr 2026 00:00:00 +0000</pubDate><description><![CDATA[<p>An MCP gateway sits between your AI agent and the MCP servers it talks to. It is the one place where routing, authentication, access control, policy, and (depending on the product) content inspection come together. Some gateways only handle routing and auth. Others also scan tool calls and responses. The term covers a range of products with different depth.</p>
<p>Without a gateway, your agent talks directly to MCP servers with no inspection. Tool descriptions can contain hidden instructions. Tool arguments can carry your API keys. Responses can inject new prompts into your agent&rsquo;s context. There is no log, no scan, no enforcement point, and no way to roll out a policy change without touching every agent.</p>]]></description></item><item><title>Generative AI Firewall: What It Is and When You Need One</title><link>https://pipelab.org/learn/generative-ai-firewall/</link><guid>https://pipelab.org/learn/generative-ai-firewall/</guid><pubDate>Tue, 14 Apr 2026 00:00:00 +0000</pubDate><description>&lt;p>A generative AI firewall, also called a GenAI firewall, is a security control that inspects traffic moving through generative AI systems and blocks the dangerous parts. Inbound prompts get scanned for jailbreaks and injection. Outbound completions get scanned for data leaks and unsafe content. Agent tool calls get scanned for exfiltration and policy violations. The firewall sits inline as a proxy, gateway, or edge service, and it only sees the traffic that reaches it.&lt;/p></description></item><item><title>AI Agent Data Loss Prevention: How DLP Works for Agents (and Where It Breaks)</title><link>https://pipelab.org/learn/ai-agent-data-loss-prevention/</link><guid>https://pipelab.org/learn/ai-agent-data-loss-prevention/</guid><pubDate>Sat, 18 Apr 2026 00:00:00 +0000</pubDate><description>&lt;p>AI agents have credentials, shell access, and the internet. When one of them gets prompt-injected (through a poisoned MCP tool description, a malicious webpage, an instruction buried in a fetched document), those credentials leave through the next HTTP request the agent makes. AI agent data loss prevention is the practice of catching the leak before it leaves the network.&lt;/p>
&lt;p>This page covers what AI agent DLP actually catches, where the leak channels live, why traditional DLP misses most of them, and the open-source self-hosted approach.&lt;/p></description></item><item><title>Claude Code Security: Harden the Agent Runtime</title><link>https://pipelab.org/learn/claude-code-security/</link><guid>https://pipelab.org/learn/claude-code-security/</guid><pubDate>Sat, 13 Jun 2026 00:00:00 +0000</pubDate><description>&lt;p>Claude Code has shell access, your API keys, and a set of MCP servers feeding content straight into its context window. That is what makes it useful. It is also the attack surface. If the agent gets tricked by a prompt injection or a poisoned MCP server, it can leak credentials, overwrite files, or run a command, and it will do all of it at machine speed before you notice.&lt;/p></description></item><item><title>EU AI Act Compliance: Article 26 Deployer Obligations and 6-Month Log Retention</title><link>https://pipelab.org/learn/eu-ai-act-compliance/</link><guid>https://pipelab.org/learn/eu-ai-act-compliance/</guid><pubDate>Sat, 07 Mar 2026 00:00:00 +0000</pubDate><description><![CDATA[<h2 id="article-266-deployer-log-retention-answer">Article 26(6) deployer log retention answer</h2>
<p>Deployers of high-risk AI systems must retain the logs the system automatically generates for <strong>at least six months</strong>, to the extent the deployer controls those logs, unless other applicable Union or national law provides otherwise. Article 12 is the design-time obligation on the provider to make the system technically capable of automatic logging. Article 26(6) is the runtime obligation on the deployer to keep the logs that result.</p>]]></description></item><item><title>Cloudflare AI Gateway: Pricing, Features, Limits &amp; Alternatives (2026)</title><link>https://pipelab.org/learn/cloudflare-ai-gateway/</link><guid>https://pipelab.org/learn/cloudflare-ai-gateway/</guid><pubDate>Tue, 05 May 2026 00:00:00 +0000</pubDate><description><![CDATA[<p><a href="/pipelock/">Pipelock</a> is the open-source agent firewall written for the network-egress boundary. Pipelock emits mediator-signed action receipts from outside the agent trust boundary, scanning HTTP, MCP, and WebSocket egress for credential leaks, prompt injection, SSRF, and tool poisoning. Cloudflare AI Gateway is something else, sitting at a different boundary: hosted at Cloudflare&rsquo;s edge, in front of LLM API calls, with caching, rate limiting, and content scanning on the prompts and completions that traverse the LLM-API surface. Both can be useful. They do not overlap as much as the marketing labels suggest.</p>]]></description></item><item><title>Pipelock Known Limitations: What It Does Not Catch</title><link>https://pipelab.org/learn/known-limitations/</link><guid>https://pipelab.org/learn/known-limitations/</guid><pubDate>Sat, 13 Jun 2026 00:00:00 +0000</pubDate><description><![CDATA[<p>Pipelock publishes what it cannot catch. A security control that claims full coverage is the one to distrust, and a reviewer needs the boundaries drawn to decide what residual risk is acceptable. This page is the standing register. It tracks the documented gaps and, per release, the false-positive and true-positive picture.</p>
<p>The authoritative engineering source is the <a href="https://github.com/pipelab-org/pipelock/blob/main/docs/bypass-resistance.md" target="_blank" rel="noopener noreferrer">bypass resistance doc</a>, which lists the evasion techniques Pipelock handles alongside the ones it does not. This page is the operator-facing summary.</p>]]></description></item><item><title>Codex CLI Security: Wrap MCP Servers and Route Egress</title><link>https://pipelab.org/learn/codex-cli-integration/</link><guid>https://pipelab.org/learn/codex-cli-integration/</guid><pubDate>Wed, 28 Oct 2026 00:00:00 +0000</pubDate><description><![CDATA[<p><a href="https://learn.chatgpt.com/docs/security" target="_blank" rel="noopener noreferrer">Codex</a> is OpenAI&rsquo;s coding agent. The <code>codex</code> CLI reads a real repository, runs commands, and calls tools from MCP servers you register with <code>codex mcp add</code>. Pipelock covers two of those paths. It wraps each MCP server with <code>pipelock mcp proxy</code>, and it inspects outbound requests from commands that honor proxy settings.</p>
<p>This page is about the Codex CLI at runtime. OpenAI&rsquo;s separate Codex Security product scans connected GitHub repositories, and OpenAI&rsquo;s own <a href="https://learn.chatgpt.com/docs/security" target="_blank" rel="noopener noreferrer">security page</a> points readers looking for it elsewhere. The full command reference for this integration lives in the <a href="https://github.com/pipelab-org/pipelock/blob/v3.8.0/docs/guides/codex.md" target="_blank" rel="noopener noreferrer">Codex integration guide</a> on GitHub.</p>]]></description></item><item><title>Pipelock v3.8 Upgrade Guide</title><link>https://pipelab.org/learn/pipelock-v380-upgrade/</link><guid>https://pipelab.org/learn/pipelock-v380-upgrade/</guid><pubDate>Mon, 09 Nov 2026 00:00:00 +0000</pubDate><description><![CDATA[<p>Install the new binary before loading configuration that uses new fields. Then move metrics scrapers, update registry references, and review the audit mode name, exceptions, presets, and repository-scan suppressions. The <a href="/blog/pipelock-v380-release/">v3.8 release notes</a> describe the changes; the <a href="https://github.com/pipelab-org/pipelock/blob/v3.8.0/docs/guides/upgrading-to-3.8.md" target="_blank" rel="noopener noreferrer">source upgrade guide</a> holds the full compatibility contract. Upgrading from an earlier line? Read the <a href="/learn/pipelock-v370-upgrade/">v3.7 upgrade guide</a> first.</p>
<h2 id="move-metrics-scrapers-to-a-dedicated-listener">Move metrics scrapers to a dedicated listener</h2>
<p>The agent-facing proxy no longer serves <code>/metrics</code> or <code>/stats</code>. Counters still collect internally, but neither route is exposed unless <code>metrics_listen</code> names a dedicated listener. Omitted, null, and blank values disable it at startup. For a scraper on the proxy host:</p>]]></description></item><item><title>Record signed receipts and verify them offline</title><link>https://pipelab.org/learn/record-verify-receipts/</link><guid>https://pipelab.org/learn/record-verify-receipts/</guid><pubDate>Thu, 05 Nov 2026 00:00:00 +0000</pubDate><description><![CDATA[<p>Pipelock can sign the record of a blocked request so you can verify it later using local files. This recipe starts from release archives, creates a disposable signing key, records one synthetic-secret block, and checks the evidence with <code>pipelock-verifier</code>. It also changes a copy and confirms that verification rejects it.</p>
<p>The <a href="/learn/what-did-my-agent-do/#try-it-on-your-own-setup">existing receipt-generation demo</a> exercises several transports from source. This recipe uses the pinned release and HTTP fetch path. <a href="/learn/agent-action-receipts/">Action receipts</a> explains the evidence format; <a href="/learn/verify-a-receipt/">Verify a receipt</a> uses published conformance fixtures instead of generating new evidence.</p>]]></description></item><item><title>Block synthetic patient data with Pipelock</title><link>https://pipelab.org/learn/block-synthetic-patient-data/</link><guid>https://pipelab.org/learn/block-synthetic-patient-data/</guid><pubDate>Thu, 05 Nov 2026 00:00:00 +0000</pubDate><description><![CDATA[<p>Pipelock can block a labeled medical record number (MRN) in a request URL when the signed <code>healthcare-phi-pii</code> bundle is loaded. Use synthetic data only. This recipe sends one planted MRN through the fetch proxy and checks the exact bundle rule that stopped it.</p>
<p>The <a href="/learn/hipaa-security-rule/">HIPAA mapping</a> explains the organizational and deployment controls this test doesn&rsquo;t cover. The <a href="/learn/community-rules/">community rules guide</a> covers signature trust, freshness and additive loading.</p>
<h2 id="starting-point">Starting point</h2>
<p>Use Ubuntu 24.04 on Linux x86-64 with Bash, curl, CA certificates and tar. The commands also use standard Ubuntu tools, including <code>grep</code>, <code>sha256sum</code> and <code>seq</code>. Start in an empty working directory with no Pipelock installation, config, installed rules or license. Run the blocks in order in the same Bash shell. This recipe uses free features.</p>]]></description></item><item><title>Block a fake secret with signed community rules</title><link>https://pipelab.org/learn/block-community-secret/</link><guid>https://pipelab.org/learn/block-community-secret/</guid><pubDate>Thu, 05 Nov 2026 00:00:00 +0000</pubDate><description><![CDATA[<p>Install the signed <code>pipelock-community</code> bundle, send a fake Buildkite-shaped token through Pipelock&rsquo;s fetch proxy, and check that the block names the community rule. The planted value isn&rsquo;t a credential and can&rsquo;t authenticate to anything.</p>
<p>The <a href="/learn/community-rules/">community rules guide</a> explains signature trust, freshness, experimental-rule filtering and updates. <a href="/pipelock-rules/">Pipelock Rules</a> describes the published bundles.</p>
<h2 id="starting-point">Starting point</h2>
<p>Use Ubuntu 24.04 on Linux x86-64 with Bash, curl, CA certificates and tar. The commands also use standard Ubuntu tools, including <code>grep</code>, <code>sha256sum</code> and <code>seq</code>. Start in an empty working directory with no Pipelock installation, config, installed rules or license. Run the blocks in order in the same Bash shell. This recipe uses free features.</p>]]></description></item><item><title>Block a poisoned MCP tool description</title><link>https://pipelab.org/learn/block-poisoned-mcp-tool/</link><guid>https://pipelab.org/learn/block-poisoned-mcp-tool/</guid><pubDate>Thu, 05 Nov 2026 00:00:00 +0000</pubDate><description><![CDATA[<p>Pipelock can refuse a Model Context Protocol (MCP) tool list before the client sees a poisoned description. This recipe compares a clean definition with a description that tells the client to ignore instructions and send a private file. The file is never opened, and the fixture makes no network requests.</p>
<p>The <a href="/learn/mcp-proxy/">MCP proxy guide</a> explains the transports and scan directions. <a href="/learn/claude-code-hooks/">Claude Code hooks</a> describes how proxy wrapping complements client-side hooks.</p>
<h2 id="starting-point">Starting point</h2>
<p>Use Ubuntu 24.04 on Linux x86-64 with Bash, curl, CA certificates and tar. The commands also use standard Ubuntu tools, including <code>grep</code> and <code>sha256sum</code>. Start in an empty working directory with no Pipelock installation, config, installed rules or license. Run the blocks in order in the same Bash shell. This recipe uses free features.</p>]]></description></item><item><title>MCP Tool Poisoning: Detection and Runtime Defense</title><link>https://pipelab.org/learn/mcp-tool-poisoning/</link><guid>https://pipelab.org/learn/mcp-tool-poisoning/</guid><pubDate>Tue, 07 Apr 2026 00:00:00 +0000</pubDate><description><![CDATA[<h2 id="mcp-tool-poisoning-in-one-paragraph">MCP tool poisoning in one paragraph</h2>
<p>MCP tool poisoning is an attack where a malicious Model Context Protocol server hides instructions inside tool metadata that an AI agent reads during discovery. The hidden instructions enter the model&rsquo;s context window as trusted documentation and the agent acts on them. Every text field in a tool schema is a potential injection surface, not just the top-level description. Rug-pulls are a sub-variant where the server returns clean descriptions to install-time scanners and poisoned descriptions after the agent is using the server. Effective defense combines pre-deploy scanning, runtime proxy inspection, session fingerprinting, and tool inventory binding.</p>]]></description></item></channel></rss>