Post-deploy discovery
Scans find risk hours or days later. An agent with write tools can cause material harm in seconds.
- Rail: pre-act intercept on tool/MCP calls
security.alphagentic.io · Guardrail Protocol v0.1
Controls that decide before the agent acts.
Tenable-class tools excel at finding exposure after the fact. Agentic systems need a different center of gravity: lanes, tools, gates, and reasoning — not only CVE lists.
Pre-act evaluation · Semantic + schema layers · Evidence ledger · Continuous adversarial evals
Pre-actGate before tool runs
ReasonedALLOW / DENY / ESCALATE + why
GraphAgent · tool · data · policy
ProofImmutable control evidence
Where exposure management falls short for agents
Respectfully: Tenable and peers are strong at asset/CVE/cloud exposure. Agentic risk is primarily action-time capability, not unpatched packages.
Wrong primary object
Hosts and CVEs dominate. Dangerous unit is the edge: agent → tool → data class → blast radius.
- Rail: action/tool graph as the control plane
Semantic bypass
Exact-match filters lose to paraphrase, encoding, and indirection — a known AI control failure mode.
- Rail: layered schema + intent + policy + human gates
Score without decision quality
Risk scores rank findings. Runtime needs “should this agent do X now?” with a traceable why.
- Rail: decision objects + reasoning traces
Non-human identity lag
Agents inherit broad credentials; toxic combinations form outside human IAM workflows.
- Rail: least-privilege lanes, no ambient admin
No closed-loop agent tests
Tickets close CVEs; few programs continuously red-team agent behavior in production paths.
- Rail: continuous adversarial eval harness as a control
Guardrail Protocol v0.1
A simple contract: every privileged act is evaluated, explained, and ledgered.
Declare lanes
Each agent gets a lane: purpose, tool allowlist, data classes, environment (sandbox/prod), human gate rules.
Intercept actions
Before tool/MCP/API execution, submit Action Intent: agent id, tool, args schema, context, urgency.
Evaluate in layers
Allowlist → lane boundary → data class → blast radius → semantic intent → optional human escalate.
Decide with reasons
ALLOW · DENY · ESCALATE. Every outcome includes structured reasons and policy version.
Ledger evidence
Immutable record for audit: who/what/when/why. Exportable for compliance — not a PDF afterthought.
Continuously eval
Adversarial suites (injection, exfil, scope creep) run like CI — guardrails that aren’t tested don’t exist.
Control Hub demo
Pick a lane and an attempted action. Rail returns a decision with reasoning layers. (Illustrative client-side demo.)
Agent lanes
Attempted actions
Recent decisions
Decision + reasoning
Smarter control model
Not “more dashboards.” A tighter loop from policy to proof.
| Layer | Tenable-class center | Rail center |
|---|---|---|
| Inventory | Assets, CVEs, cloud resources | Agents, lanes, tools/MCP, data classes, policies |
| Time | Scan / expose / remediate cycles | Milliseconds before action + continuous eval |
| Decision | Risk score & priority queue | ALLOW / DENY / ESCALATE + structured why |
| Identity | Human-first IAM + some NHI signals | Non-human lanes as first-class least privilege |
| Proof | Findings & reports | Immutable action ledger & policy versioning |
| AI abuse | Emerging posture / exposure views | Semantic + schema guardrails on the wire |
Evaluate endpoint
POST /v1/evaluate
Submit intent; receive decision + reasons + ledger id. Wire into agent runtimes and MCP proxies.
Versioned controls
Policies ship as signed packs. Roll forward/back. Every decision cites pack version.
Audit export
One stream for SecOps, GRC, and incident review — not three tools stitched by screenshots.
Build path
Honest scope. Ship the control plane before the marketing fleet.
Protocol + hub demo
This site. Decision model. Gap thesis. Design system aligned to AlphAgentic craft.
Evaluate API + proxy
Server-side policy engine, MCP/tool proxy, signed ledger, sandbox agent fixtures.
Enterprise connectors
IdP, SIEM export, ticketing, continuous eval CI, multi-tenant control hub.
Guardrails with reasoning. Controls with proof.
security.alphagentic.io — AlphAgentic Rail
Talk to us