Network-level security for autonomous agents.
Field notes on prompt injection, data exfiltration, MCP gateways, AI governance, and what actually happens on the wire when an LLM acts on its own.
Agent G vs Lakera Guard: Prompt Classifier vs Egress Enforcement
Lakera Guard classifies prompts and outputs before they reach the model. Agent G enforces what the agent actually does on the wire. Here is how the two layers differ and why serious agent deployments need action-level enforcement, not just detection.
Read article →
A2A (Agent-to-Agent) Traffic: The Next Unmonitored Egress Surface
Agent-to-agent (A2A) traffic is a fast-growing egress surface that most guardrails never inspect. Here is what A2A calls look like on the wire and how to bring them under egress enforcement.
Read article →
NeMo Guardrails and Guardrails AI vs an Egress Firewall
In-process dialogue and validation rails run inside the agent process, so a compromised agent can bypass them entirely. Here is why the egress boundary is the enforcement layer NeMo Guardrails and Guardrails AI cannot reach.
Read article →
How to Prevent Data Exfiltration via Markdown Image Rendering in LLM Apps
Markdown image rendering is a silent exfiltration channel in LLM apps. Learn how attackers smuggle data out through image URLs, and how Agent G blocks the outbound fetch at the egress boundary.
Read article →
The Business Case for an AI Agent Firewall
A CFO-ready framework for quantifying the cost of an ai agent breach and justifying inline egress enforcement, human-in-the-loop approval, and wire-level audit logging.
Read article →
Egress Control for AI Agents on Azure: Locking Down the VNet
A deployment guide for enforcing egress control on AI agents running in Azure, covering NSGs, Azure Firewall blind spots, IMDS lockdown, and where an inline agent firewall like Agent G fits in the VNet.
Read article →
Shadow MCP: Finding the Servers Your Agents Might Be Calling
Shadow MCP servers are the unsanctioned endpoints your agents call without approval. Here is how to surface them from egress logs and gate them at the proxy.
Read article →
Deep Tool-Argument Inspection: The MCP Gateway Blind Spot
Most MCP gateways route and authenticate tool calls but never read the arguments or responses. That is the blind spot attackers use. Here is how deep tool-argument inspection closes it at the egress boundary.
Read article →
MCP 'Connection Closed' Error: Causes and Fix
The MCP connection closed error usually means the transport died before the handshake completed. Here are the concrete causes, a diagnostic checklist, and how egress policy makes MCP connections both reliable and auditable.
Read article →
How to Prevent MCP Tool Poisoning at Runtime
Static review cannot catch an MCP tool that changes behavior after approval. Here is how to prevent MCP tool poisoning at runtime by constraining what tool calls do on the wire.
Read article →
Cloudflare AI Gateway + an Egress Proxy: The Combined Agent Stack
Cloudflare AI Gateway sits in front of your LLM calls for caching, analytics, and rate limiting. It does not inspect or block the outbound tool actions an agent takes. Here is how the two layers combine into one coherent agent security stack.
Read article →
Stopping Agent Data Exfiltration via Email, Slack & Webhooks
Legitimate messaging and webhook tools are the easiest path for an autonomous agent to leak data. Here is how the exfiltration happens on the wire and how Agent G inspects and blocks those outbound calls.
Read article →
SSRF in AI Agents: Why a Tool Call Is All It Takes
A single agent tool call can trigger server side request forgery against your cloud metadata service. Here is how SSRF in AI agents works and how to block it at the egress boundary.
Read article →
How to Secure MCP Server Authentication and OAuth
MCP authentication and OAuth pitfalls like token passthrough and over-scoped tokens quietly widen your agent attack surface. Here is how to harden them and enforce identity and egress around every authenticated MCP call.
Read article →
Deep Tool-Argument Inspection: The MCP Gateway Blind Spot
Most MCP gateways route and authenticate tool calls but never read the arguments or responses. That blind spot is where exfiltration and destructive actions slip through. Here is how deep tool-argument inspection closes it.
Read article →
LiteLLM Proxy vs an AI Agent Firewall: Where LiteLLM Proxy Security Stops
LiteLLM is a great LLM gateway for routing, keys, and cost, but it does not inspect or block the outbound tool actions an agent takes. Here is where LiteLLM proxy security ends and an AI agent firewall begins.
Read article →
SSRF in AI Agents: Why a Tool Call Is All It Takes (and How to Block It at Egress)
A single agent tool call can trigger server side request forgery and steal cloud credentials. Here is how SSRF works in AI agents and how to block it at the egress boundary.
Read article →
How to Evaluate an AI Agent Firewall: RFP Checklist
A buyer-enablement checklist covering the must-have egress, human-in-the-loop, policy-as-code, and logging capabilities you should demand from any AI agent firewall before you sign.
Read article →
DNS as the Agent Exfil Channel Nobody Monitors
Autonomous agents can leak secrets one DNS query at a time, right past your allowlists and TLS inspection. Here is how DNS exfiltration works and how to block it at egress.
Read article →
HIPAA Compliance for AI Agents Handling PHI
A practical guide to HIPAA compliance for AI agents that handle PHI, mapping the Security Rule to concrete egress controls that contain data, enforce minimum necessary, and produce audit evidence.
Read article →
Inference-Boundary vs Network-Boundary AI Security: Why You Need Both
Prompt classifiers guard the inference boundary; egress enforcement guards the network boundary. Here is why serious agent security needs both layers, and what each one can and cannot stop.
Read article →
The Platform Engineer's Guide to Agent Egress
A practical guide for platform and DevSecOps teams operationalizing agent egress as shared infrastructure, from proxy placement to policy-as-code and audit.
Read article →
Agent Action Receipts: Verifiable Audit Evidence From Outside the Trust Boundary
Agent action receipts turn raw egress telemetry into signed, verifiable audit evidence captured outside the agent's own trust boundary. This guide explains how tamper-proof receipts work, why in-process logs fail compliance, and how Agent G generates them at the network edge.
Read article →
Protect AI Alternative: Model Scanning vs Egress Enforcement
Protect AI scans models and surfaces posture risk, but it never blocks the outbound action your agent takes at runtime. Here is how Agent G enforces egress on the wire, and why you likely need both layers, not one.
Read article →
Prevent Destructive AI Agent Actions: Blocking rm -rf, DROP TABLE, and Rogue Wire Transfers
Autonomous agents now hold shell access, database credentials, and payment APIs: one bad token prediction away from an irreversible command. Here is how to block destructive AI agent actions at the network boundary, before they ever execute.
Read article →
Engineering Human-in-the-Loop Approval for Risky Agent Actions
Autonomous agents need a human checkpoint before they take irreversible actions, but a clumsy gate destroys throughput. Here is how to engineer a fast, auditable human-in-the-loop approval workflow with Agent G.
Read article →
Build vs Buy: An AI Agent Egress Firewall
Building a DIY egress proxy from Squid, Envoy, and OPA looks cheap until you price the ongoing maintenance, coverage gaps, and on-call burden. Here is an honest build-vs-buy framework for platform teams scoping an AI agent egress firewall.
Read article →
Securing LangGraph Agents in Production
LangGraph's stateful graphs give agents memory, loops, and node-level tool calls, and a wider blast radius. This guide shows how to harden LangGraph agents for production and enforce egress and human approval on every node-level action.
Read article →
Engineering Human-in-the-Loop Approval for Risky Agent Actions
Autonomous agents will eventually try to do something irreversible. The fix is a human-in-the-loop approval gate that intercepts risky actions at the wire, without grinding your pipeline to a halt. Here is how to engineer it.
Read article →
Securing AI Coding Agents: Cursor, Claude Code, and Devin at the Egress Boundary
AI coding agents run with shell access, live credentials, and network reach inside your CI. When a repo or prompt turns hostile, the only reliable backstop is egress control. Here's how to lock it down.
Read article →
Default-Deny Egress: Building an Allowlist Your LLM Can't Talk Around
Autonomous agents decide their own outbound calls at runtime, which means a denylist will always be one step behind. This playbook shows how to build a default-deny egress allowlist your LLM literally cannot talk around, and how Agent G enforces it as a drop-in proxy.
Read article →
AWS Bedrock Guardrails vs. an AI Agent Firewall
Bedrock Guardrails work at the model boundary; an agent firewall works on the wire. Where each fits, what each misses, and why production agents need both layers.
Read article →
What Actually Happens on the Wire When an LLM Agent Acts
The reasoning is invisible, but every action an autonomous agent takes becomes bytes on the wire. Here's what that traffic looks like and why the network boundary is where you enforce it.
Read article →
Browser Agent Security Risks: What Breaks With AI in a Browser
Browser-using agents expand the blast radius. Here's how prompt injection, credential leakage, and CSRF-style attacks change shape and what the network layer can do about them.
Read article →
Preventing LLM Data Exfiltration: A Network-Layer Playbook
Data exfiltration meaning, in agent terms: any outbound call that sends data somewhere it shouldn't. We break down the egress patterns and how to block them.
Read article →
Demystifying MCP Gateways: When You Need One and What It Does
An MCP gateway sits between your agent and the tool servers it talks to. What it does, what it doesn't, and where guardrails belong.
Read article →
Claude 529 Overloaded Error: Causes and Safe Resolution
What the Claude overloaded error actually means, why it surfaces in agent loops, and how to harden retries without amplifying load.
Read article →
AI Governance Auditing Framework for Agent Workloads
Mapping NIST AI RMF and ISO 42001 controls to agent traffic: what to log, what to block, and what to escalate.
Read article →