Research investigation R1118 / comparison

Codex and Claude Code command allowlists do not cover separately launched MCP servers

Both products document destination controls for sandboxed commands and descendants, but network access, domain enforcement and administrator restrictions are separate configuration decisions. Neither command allowlist extends to separately launched MCP-server networking.

Current public editionNov 19, 2026Nov 19, 2026
Verified observations
0

0 measured fields

Supported claims
9

9 material findings

Cited sources
6

6 primary or authoritative

Research score
78

Automated topic and evidence score

Interactive figureCodex and Claude Code command allowlists do not...
CSV JSON
Data statusAwaiting verified observations

The sample covers the supplied local-session documentation for Codex and Claude Code, not every deployment or execution path. Codex documents sandboxed scripts, programs and child processes as covered; Claude documents shell commands and descendants as covered. The shared MCP finding concerns separately launched servers, not every possible MCP deployment. Sources are citation-only.

Verified observationNo chart values are being inferredLast updated Nov 19, 2026
Coverage note

The sample covers the supplied local-session documentation for Codex and Claude Code, not every deployment or execution path. Codex documents sandboxed scripts, programs and child processes as covered; Claude documents shell commands and descendants as covered. The shared MCP finding concerns separately launched servers, not every possible MCP deployment. Sources are citation-only.

Dataset ID
spd:codex-vs-claude-code-when-does-a-network-allowlist-actually-constrain-commands-879a935f
Stable URL
/research/codex-vs-claude-code-when-does-a-network-allowlist-actually-constrain-commands-879a935f
Version
Live
Coverage
Live collection
Records
0
Fields
7
Updated

Read the data

The records behind the figure

CSV JSON
Codex and Claude Code command allowlists do not cover separately launched MCP servers data records
EntityMetricValueUnitObservedSourceTransform

Measurement technique

How to read this report

  1. 01Synthesize the saved evidence without treating its recovery as a new collection or experiment. Preserve the November 18, 2026, 21:31 UTC cutoff.
  2. 02Use the six supplied primary documents: OpenAI security, managed configuration and configuration reference; Anthropic sandboxing, settings precedence and settings reference. The candidate proposed five documents, but the supplied evidence includes a sixth settings-reference document.
  3. 03Plan an evidence_matrix with columns for product, local-session scenario, required settings, shell and descendant coverage, exceptions, administrator control, platform/version caveat and supporting document.
  4. 04Code matched scenarios as documented constrained, approval-dependent, outside scope or unspecified. Keep documented configuration behavior separate from observed runtime behavior.
  5. 05Populate rows for networking disabled or enabled, destination enforcement, managed-only domains, escalation or exclusions, replacement proxies and separately launched MCP servers. Do not turn these categorical findings into a quantitative chart.
Next report / 01AI Model Economics Index All research reports
YOUR READING SPACE

Notifications