section // projects46 notes8 research4 cvelast sync 2026-09-01
← Open-source projects
Agent security · standalone project preview

See what an agent can actually reach.

AgentMordor turns scattered tools, identities, data access, and approval rules into one evidence-backed map. Follow the path from untrusted input to a real-world action before an agent reaches production.

View the interface conceptRead the product scope
01No execution on the blogThis page only presents the standalone product.
02Evidence per connectionEvery line points back to its source.
03Change-awareNew authority is visible in every pull request.
Promotional preview

This blog page does not scan configurations, import reports, or run the AgentMordor CLI. The interface below uses fixed synthetic data only to explain the standalone product.

Interface concept

One workspace. Three readable answers.

Select an access path, inspect its evidence, then switch to Changes or Controls. This promotional preview uses fixed synthetic data and never reads a local file or touches a real environment.

AMAgentMordorUI concept
workspaceacme-security / agent-platform
synthetic previewno scan
Effective access mapDeclared configuration + inferred reach
InputAgentToolIdentityAction
Input
Agent
Tool
Identity / approval
Reachable action
GitHubIssue commentexternal input
Automaticrelease-triagetrigger: opened
MCP toolcloud.deploywrite operation
AWS roleprod-deployerapproval: none
ProductionDeploy servicechanges workload
GitPull requestrepository content
Automaticcode-reviewtrigger: synchronize
Local toolshell.execarguments: any
CI tokenservice-cishared identity
Secret storeRead production12 secret paths
ConsoleAnalyst requestauthenticated user
Manualresearch-agentsandboxed
Bound toolreport.writefixed destination
Required gateSecurity reviewernamed approval
ArchivePublish reportwrite only
fixed synthetic data · 0 processes executedpromotional UI concept

Product functionality lives in the separate AgentMordor project, not in this blog.

Product boundary

Built around the question scanners usually skip.

AgentMordor does not guess whether a prompt sounds dangerous. It follows configured access until it reaches a file, secret, API, cloud role, or real-world action—and shows the evidence used for every connection.

01

Discover

Start with explicit MCP JSON inputs, then add host-specific adapters without silently searching the operator's home directory.

v0.1 · explicit MCP JSON input
02

Connect

Turn isolated permissions into readable end-to-end paths, including where input comes from and what can change.

Input → agent → identity → action
03

Explain

Attach a source file, line, confidence, and plain-language reason to every inferred connection.

No unexplained risk scores
04

Diff

Store a safe baseline and reveal only the new authority introduced by a pull request or configuration change.

JSON · SARIF · HTML
Build sequence

Small core. Serious evidence.

  1. v0.1
    Local inventory

    Read-only MCP JSON parser, deterministic findings, text/JSON reports, redaction tests, and synthetic fixtures.

    Now
  2. v0.2
    Authority diff

    Baseline comparison, policy thresholds, SARIF output, CI annotations.

    Next
  3. v0.3
    Verified controls

    Safe deny tests, approval validation, framework mappings, signed evidence.

    Later