Guard research desk
Decide with evidence, not vendor claims.
Neutral decision guides and comparisons for AI agent security across model, tool, runtime, and supply-chain boundaries. Each guide states its direct answer, honest limitations, and the sources it was reviewed against — with a visible review expiry.
6 guides3 comparisonsNamed facts expire with each review window
Decision guides
- GuideAI Agent Security LayersAI agent security is layered across the model, harness and orchestration, tools and execution environment, supply chain, and adjacent identity, secret-management, endpoint, network, DLP, and evidence controls. No single layer is sufficient. For coding agents, consequential-action controls belong at supported harness or execution boundaries, while supply-chain controls evaluate packages and artifacts before installation. Evaluate coverage by exact surface, decision timing, and failure behavior rather than product category labels.Reviewed 2026-09-13Read
- GuideMCP Scanning vs Runtime EnforcementMCP scanning and runtime enforcement address different points in the MCP server lifecycle. Scanning evaluates a server definition before it is registered, checking for known patterns, permissions, and configuration risks. Runtime enforcement intercepts tool calls after the server is registered, requiring approval before execution. Scanning is preventive; enforcement is reactive. Both are needed: scanning catches configuration issues, enforcement catches behavioral risks. HOL Guard provides runtime enforcement, not scanning.Reviewed 2026-08-09Read
- GuideRuntime Guardrails and Supply Chain SecurityRuntime guardrails and supply chain security tools address different phases of the software lifecycle. Supply chain tools verify packages at install time, checking for known vulnerabilities, typosquatting, and malicious code. Runtime guardrails intercept the agent when it tries to install, read, or execute something. Supply chain security is preventive (before install); runtime guardrails are reactive (during execution). Both are needed because a verified package can still be used in a risky way, and a runtime guardrail cannot inspect every package before install.Reviewed 2026-09-13Read
- GuideProtect Secrets from AI Coding AgentsAI coding agents can read secrets (.env files, SSH keys, API tokens) if they have file system access. Protection approaches include: environment variable isolation (agents run without secrets in scope), runtime guardrails (intercept file reads before execution), and secret management tools (vaults that agents never touch directly). Runtime guardrails like HOL Guard intercept attempts to read .env, .npmrc, and SSH key files. Environment isolation prevents secrets from being accessible at all. Both approaches are complementary.Reviewed 2026-09-15Read
- GuideSecure MCP for AI Coding AgentsSecuring MCP across AI coding agents requires per-client configuration review plus controls on the action surfaces each client actually exposes. MCP standardizes tools and context, but Codex, Claude Code, Cursor, Antigravity CLI, and OpenCode configure and approve servers differently. HOL Guard coverage is harness- and event-specific: on supported surfaces it can review MCP configuration changes or consequential MCP tool actions, while other paths are preflight or configuration review only. Combine the client’s native controls with least-privilege server configuration and Guard only where its current support contract applies.Reviewed 2026-09-13Read
- GuidePrompt Injection ProtectionHOL Guard can reduce impact when influenced reasoning reaches a supported consequential action boundary. On those surfaces it can pause, block, or ask before a supported shell, file, MCP, skill, or package action executes. It does not scrub model context and is not a complete prompt-injection preventer. Model guardrails alone are insufficient. Coverage is harness- and event-specific. A scan is not a safety guarantee. Guard is not a WAF, EDR, or MDM. HOL makes Guard.Reviewed 2026-09-13Read
Comparisons
- ComparisonRuntime Guardrails vs Native Agent ControlsNative agent controls and external runtime guardrails both constrain agent actions, but at different boundaries. Native controls are built into each harness and vary by mode and version. HOL Guard adds a separate local policy boundary on the specific harness and event surfaces it supports. Neither replaces the other. Guard Gauntlet is the live method, and the July 2026 fixture benchmark must not be read as a Gauntlet score or as live attack-resistance evidence.Reviewed 2026-09-13Read
- ComparisonAI Coding Agent Security Tools ComparisonAI coding agent security tools fall into three categories: runtime guardrails (HOL Guard), static analysis scanners (Snyk Code, GitHub Advanced Security), and IDE-integrated security (built-in harness controls). Each tool type addresses a different layer. Runtime guardrails intercept actions before execution. Static analysis scans code for vulnerabilities. IDE controls provide per-harness permission prompts. No single tool covers all layers. Compare tools on identical criteria: what they intercept, when they act, what they cover, and what they miss.Reviewed 2026-09-13Read
- ComparisonBest AI Agent Security PlatformsThere is no universal AI agent security platform. The right control depends on the boundary: model traffic, MCP and tool mediation, artifacts, native harness permissions, or runtime actions on a developer machine. Compare these as complementary layers on identical criteria. HOL Guard is local-first runtime policy for supported AI coding-agent action surfaces. It is not a WAF, EDR, MDM, secrets manager, cloud MCP gateway, or universal prompt filter. A scan is not a safety guarantee.Reviewed 2026-08-09Read
Go deeper
Continue into the security hub
Threat explainers, evidence dossiers, and advisories back every recommendation these guides make.