Skip to content

Agent Security Tooling Landscape

This is the living landscape — we keep it current as the space moves. For point-in-time strategic reads (“what we thought, when”), see the dated snapshots: the most recent is May 2026, preceded by April 2026.

Last reviewed: 2026-05-23


The agent security space has matured past “can we intercept the agent?” into “can we prove what it did, in a form a third party will accept?” Every enforcement approach — MCP proxying, egress firewalling, kernel-level enforcement, application-level policy engines, and enterprise gateways — has at least one serious implementation. Microsoft’s Agent Governance Toolkit (AGT, v3.0 Public Preview) is the most comprehensive single project; it now integrates directly into the Agent Framework, making governance a framework default rather than an opt-in install.

The decisive shift in 2026 is provenance: signed, hash-chained receipts went from a differentiator to table stakes. What now separates projects is the trust boundary — whether the process that signs the audit trail is the same process being audited — and whether the record is portable enough for an external verifier.

The space segments into five layers:

LayerWhat it doesKey players
MCP Gateways (commercial)Managed proxy for MCP traffic with auth, audit, rate limitingMintMCP, Peta, TrueFoundry, Lasso, Gravitee, Traefik Hub
Agent Firewalls (open-source)Intercept + scan agent traffic (MCP and/or HTTP)Pipelock, mcp-firewall (ressl), Agent Wall, mcp-firewall (dzervas)
Kernel-Level EnforcementOS-level syscall interception, sandboxingagentsh / Canyon Road, Anthropic sandbox-runtime
Governance FrameworksApplication-level policy engine, identity, complianceMicrosoft AGT, GitHub Agentic Workflows
Audit & ProvenanceThe provenance record is the productAgent Receipts, nono, Asqav, InALign

The category that crystallized in May 2026: tools whose primary product is the provenance record rather than enforcement. The differentiating axis is where the signing key lives (inside vs. outside the audited process) and how portable the record is.

Primary productSigning locusEnvelopeNotes
Agent ReceiptsProvenance recordOut-of-agent daemonW3C VC (Ed25519, RFC 8785)Keys + chain outside the audited process; one chain across all channels
nonoProvenance recordOut-of-process (append-only Merkle log)Sigstore/Rekor anchoredClosest on the boundary axis; binary-identity binding
AsqavProvenance recordIn-agent / SDKSigned records (ML-DSA-65, RFC 3161 timestamps)Audit-first; post-quantum signing, EU AI Act audit packs
InALignAlignment + auditIn-processSHA-256 hash chainMCP-native; MITRE-mapped detection, EU AI Act checks

Several of these entrants are weeks old; rows carry verified specifics where available, not fabricated detail. The axis that matters is in-agent vs. out-of-agent signing and record portability — and on the first question, nono lands on the same side as Agent Receipts, so the differentiation runs through the portable W3C VC envelope and the single cross-channel chain. See the audit boundary belongs outside the agent for why the boundary is the durable property.


Detailed Comparison: Open-Source Agent Security Tools

Section titled “Detailed Comparison: Open-Source Agent Security Tools”

These are the tools most relevant to an individual builder or small team entering the space.

Microsoft AGTPipelockmcp-firewall (ressl)agentsh (Canyon Road)Agent Wallmcp-firewall (dzervas)
Releasedv3.0 Public Preview (Apr 2026)Jan 2026Feb 20262025–2026Feb 20262026
LanguagePython (primary), TS, .NET, Rust, Go SDKsGoPythonGo + system-levelNode.jsRust
LicenseMITApache 2.0AGPL-3.0 (commercial available)Source-available (commercial)MITMIT
ApproachApplication middleware + Agent Framework integrationEgress proxy + MCP proxyMCP stdio proxy + SDK libraryKernel enforcement (Landlock, FUSE, ptrace, seccomp)MCP stdio proxyClaude Code pre-tool-use hook
MCP proxyNo (framework adapters)Yes (stdio)Yes (stdio)No (syscall-level)Yes (stdio)No (hook-based)
HTTP/egress proxyNoYes (7-layer scanner)NoYes (network proxy)NoNo
Shell/command controlNoNoNoYes (shell shim, ptrace)NoNo
File I/O controlNoIntegrity monitoringNoYes (FUSE, Landlock)NoNo
Policy engineYAML + OPA/Rego + CedarYAML configYAML + OPA/RegoYAML policyYAML configJsonnet
DLP / secret scanningNoYes (regex, entropy, env leak)Yes (response scanning)Yes (output redaction)YesNo
Prompt injection detectionMCP scanner moduleYes (response scanning)Yes (8 inbound checks)NoYesNo
Cryptographic identityEd25519 DIDs + ML-DSA-65Ed25519 signingEd25519 audit chainNoNoNo
Audit loggingStructured + OTELJSON + Prometheus; EvidenceReceipt v2JSON + signed hash chainStructured + OTELJSONNo
Compliance reportingEU AI Act, NIST, HIPAA, SOC 2, OWASPOWASP mappingDORA, FINMA, SOC 2NoNoNo
DashboardNoPrometheus/stats endpointYes (web UI)Via Watchtower (commercial)Yes (web UI)No
Framework integrationsAgent Framework + 12+ (LangChain, CrewAI, AutoGen, etc.)Claude Code, CursorClaude Desktop, Cursor, any MCP clientVercel, E2B, Daytona, Cloudflare, etc.Any MCP clientClaude Code, Copilot CLI
Trust modelSame-process middlewareCapability separation (proxy has no secrets)Same-process proxyKernel-enforced isolationSame-process proxyHook-based

Detailed Comparison: Commercial MCP Gateways

Section titled “Detailed Comparison: Commercial MCP Gateways”
MintMCPPeta (Agent Vault)TrueFoundryLasso SecurityGraviteeTraefik Hub
TypeManaged SaaSCredential vault + gatewayAI platform + gatewaySecurity platform + OSS gatewayAPI gateway + MCPReverse proxy + MCP middleware
OSS componentLLM Proxy (partial)NoNoYes (mcp-gateway, Apache 2.0)NoNo
Key differentiatorOne-click deploy, pre-built connectorsZero-trust vault, agents never see raw keysLow latency (3–4ms), 350+ rpsPlugin-based guardrails, PII detection (Presidio)Protocol-aware, method-level governanceExtends existing Traefik deployments
Auth modelOAuth 2.0Scoped time-limited tokensOAuth 2.0 OBOAPI key + pluginsStandard API gateway authStandard Traefik auth
Human-in-the-loopNoYes (approval workflows)NoNoNoNo
ComplianceSOC 2SOC 2VariesGartner Cool Vendor 2024Enterprise certificationsEnterprise certifications
Best forFast deployment, non-security-specialist teamsRegulated industries, credential managementHigh-throughput, perf-sensitive deploymentsSecurity-first orgs wanting OSS flexibilityOrgs already on GraviteeOrgs already on Traefik

GitHub Agentic WorkflowsCloudflare Enterprise MCPGitHub Copilot Agent Firewall
ScopeFull defense-in-depth for GH Actions agentsWAF + AI Gateway for MCP serversDomain allowlist for Copilot cloud agent
ArchitectureKernel isolation + MCP gateway + integrity filteringWAF in front of MCP, portal pattern (N servers → 2 tools)iptables-based egress filtering
Policy modelDeclarative YAML (network, integrity levels)WAF rules + AI Gateway configDomain allowlist (org or repo level)
LimitationsGitHub Actions onlyCloudflare stack onlyOnly covers agent-started processes, not MCP servers; bypassable
NotableTrust-scored content filtering (merged/approved/unapproved)Published April 15, 2026 — their own internal deploymentHonest about limitations in their own docs

ApproachEnforcement guaranteeBypass riskSetup complexityCoverage scope
Kernel-level (agentsh)Strongest — syscall interceptionVery low (requires kernel exploit)High (kernel 6.7+, FUSE, capabilities)Shell, filesystem, network, processes
Egress proxy (Pipelock)Strong for network — capability separationMedium (agent could use alternative channels)Low (single binary)Network egress, MCP responses
MCP stdio proxy (mcp-firewall, Agent Wall)Moderate — protocol-level interceptionMedium (only covers MCP channel)Low (wrap command)MCP tool calls and responses only
Application middleware (Microsoft AGT)Weakest — same trust boundary as agentHigh (agent can bypass if compromised)Low (pip install)Whatever the framework exposes
Hook-based (dzervas/mcp-firewall)Moderate — pre-execution checkMedium (depends on client enforcement)Very lowTool calls in supported clients

Audit & provenance tools are not enforcement tools — they record rather than block. The analogous axis for them is the trust boundary: an in-process signer shares the agent’s blast radius; an out-of-agent signer does not.


Multiple projects have independently converged on the same cryptographic and protocol primitives. By May 2026, signed hash-chained receipts are no longer differentiating — they are table stakes.

PrimitiveUsed by
Ed25519 signingMicrosoft AGT, Pipelock, mcp-firewall (ressl), Asqav, nono, Agent Receipts
SHA-256 hash chainingmcp-firewall (ressl), Agent Receipts
DIDs (Decentralized Identifiers)Microsoft AGT, Agent Receipts
OPA/Rego policiesMicrosoft AGT, mcp-firewall (ressl)
Cedar policiesMicrosoft AGT
Signed, hash-chained receipt (any scheme — table stakes)Agent Receipts, Pipelock (EvidenceReceipt v2), Asqav, nono, InALign, Microsoft AGT
W3C Verifiable Credentials (data model envelope)Agent Receipts only — competitors sign receipts with other schemes (Asqav: ML-DSA-65; nono: Sigstore/Rekor; Pipelock: Ed25519 JSONL), not the VC envelope
Sigstore / Rekor transparency lognono
OWASP Agentic AI Top 10Microsoft AGT, Pipelock, mcp-firewall (ressl)
YAML policy configAll enforcement projects
Out-of-agent signing / logAgent Receipts (daemon owns keys + chain); nono (append-only log the agent can’t reach); others sign in-process
Hook-based emissionClaude Code (full tool-surface coverage), Codex CLI (partial — no WebSearch, partial shell); Agent Receipts aligns with both hook surfaces

EU AI Act Article 12 requires automatic, tamper-evident logging of high-risk AI system operation over its lifetime — nearly verbatim, the problem a hash-chained, independently-signed receipt solves. August 2, 2026 is a genuine AI Act milestone, though some Annex III high-risk obligations were reportedly deferred (to late 2027) in a recent political agreement, so the deadline is softer than a single date implies. The direction of travel — tamper-evident logging as a legal expectation — is the durable signal.

EffortVenueWhat it isMaturity
EvidenceReceipt v2PipelockReceipt format (Ed25519, evidence.jsonl); bid to standardize the firewall-emitted recordShipping
W3C VC Data IntegrityW3CThe envelope our receipts use; CG argues canonicalization (JSON-LD vs. RFC 8785)Recommendation
IETF draft-sharifIETFdraft-sharif-agent-audit-trail, a logging format for agent actionsIndividual I-D (not WG-adopted)

General supply-chain transparency work (IETF SCITT, COSE receipts) is the substrate some of these build on, but it is software-supply-chain-oriented, not agent-audit-specific, so we don’t count it as part of this race.

Agent Receipts’ position: the receipt is a W3C VC; the differentiator is the out-of-agent boundary plus one chain across channels; the format should be portable, and we will align with a credible standard rather than fork one.


Areas that remain underserved despite the crowded landscape:

GapDescriptionWho’s closest
Audit independent of the audited processSigning keys/log held outside the agent so the trail survives a compromised agentAgent Receipts (daemon process separation) and nono (append-only log the agent can’t reach); most other audit-first tools sign in-process
Unified cross-channel auditCorrelating MCP calls + REST calls + shell commands + browser actions into one timeline per agent sessionAgent Receipts (one daemon, one chain across channels); Canyon Road Watchtower (commercial/closed)
CISO-ready reportingPDF/HTML reports a security team can review to approve agentic AI adoptionmcp-firewall (ressl) has compliance reports; Microsoft AGT has framework mappings; neither produces turnkey CISO artifacts
HTTP/OpenAPI interceptionPolicy-enforced proxy for agent REST API calls (not just MCP)Pipelock (egress proxy); agentsh (network proxy) — but neither is OpenAPI-schema-aware
Browser automation governanceIntercepting Puppeteer/Playwright/CDP actions with policy enforcementNobody (GitHub Copilot firewall explicitly doesn’t cover this)
Policy portability standardA way to express agent policies that works across toolsMicrosoft AGT supports 3 languages but no cross-tool standard exists
Agent identity federationVerifying agent identity across organizational boundariesMicrosoft AGT (SPIFFE/SVID) is closest; still early

This document is kept current. For the point-in-time strategic read, see the May 2026 snapshot.