← Back to Axiom
AXIOM / SECURITY LAB

Pentesting.Evidence you can inspect.

What was tested. What the guard refused. What remains unverified. Explore the project's published lab results, inspect the application captures and take the documentation with you.

Controlled Windows VM · Synthetic payloads · Project-reported evidence

10harnesses in the targeted sweep
15vectors per swept harness
13documented probe batches
Windows 11isolated lab environment

The decisions.
Without the guesswork.

Counts verified against the saved raw event log and the historical chart capture. View the same 3,189 blocked records by harness or by reason. These are repeated guard decisions, not unique threats or an overall detection-rate benchmark.

Recorded guard blocks

Accumulated lab snapshot / GUARD_BLOCKED

  1. Claude Code844
  2. OpenCode335
  3. GitHub Copilot238
  4. Codex237
  5. Cursor225
  6. Antigravity225
  7. Gemini223
  8. ZCode223
  9. DeepSeek223
  10. Cline208
  11. Windsurf208
Source: verified raw event snapshot · 11 harness / 10 reason categoriesView source capture   /   Download chart data (CSV)

Two datasets, two scopes. The accumulated chart includes 11 harness categories. The separate targeted sweep below covers 10 named harnesses with 15 vectors each. The chart counts must not be presented as the sweep's success rate.

Inspect the targeted harness sweep — 10 harnesses × 15 vectors
Latest complete saved suite (sweep-072.txt): 10 harnesses × 15 cases, with 13 denials and 2 allowed controls per harness.
HarnessHookVectorsMatched expectations
Claude Codeaxiom-claude-hook.py1513 denied + 2 allowed
Codexaxiom-codex-hook.py1513 denied + 2 allowed
Gemini CLIaxiom-gemini-hook.py1513 denied + 2 allowed
Clineaxiom-cline-hook.py1513 denied + 2 allowed
GitHub Copilot CLIaxiom-copilot-hook.py1513 denied + 2 allowed
Cursoraxiom-cursor-hook.py1513 denied + 2 allowed
Antigravityaxiom-antigravity-hook.py1513 denied + 2 allowed
Windsurfaxiom-windsurf-hook.py1513 denied + 2 allowed
ZCodeaxiom-zcode-hook.py1513 denied + 2 allowed
DeepSeek / DSHaxiom-deepseek-hook.py1513 denied + 2 allowed

Provisioned hook replays; several editor hosts were not installed. Sweep 067 is incomplete, sweep 073 is absent, and later raw headers retain build 0.59.0. The detailed report and CSV appendices preserve these limitations.

See the application.
Inspect the evidence.

Faithful application-window and content crops, plus two owner-supplied graph captures. Desktop and unrelated window chrome have been removed. Captions retain the snapshot context and enforcement limits; in-app overlays remain where removing them would alter evidence.

Read it. Download it.
Keep the context.

Read the detailed source-based report, inspect the cases and files, or take the complete public documentation bundle. Every format preserves the scope and limitations of these project-reported results.

Prepared 17 November 2026 from historical project records, with private information excluded. This is not an independent audit or certification. Credentials, raw sensitive inputs and executable samples are not distributed. Screenshots open as full-resolution application/content crops.

Additional data: Harness matrix · 1,785 sweep observations · Hardening gaps · Canary inputs · Report in Markdown

Test the boundary.
State the limit.

The published workflow installs Axiom in an isolated Windows 11 VM and drives each agent hook with synthetic payloads. A gap leads to a fix, a regression test and a re-test of the installed build.

01 / CONTROLLED ENVIRONMENT

An installed build.

Record the Axiom version, operating system, active policy, hook and knowledge state. The captures span several builds, including 0.91.0, 0.93.0 and 0.94.0.

02 / SYNTHETIC PROBES

Explicit attack families.

Protected reads, credential stores, command execution, persistence, exfiltration, tampering and prompt injection are exercised through documented agent harnesses.

03 / OUTCOMES & RECEIPTS

Separate kinds of proof.

Observation, guard refusal and an independently verified OS block are different results. A ruleset count or a threat graph does not prove a host-level action was denied.

Explore the 13 probe batches
  • Batch 01–02
    UNC/admin-share reads; findstr, grep, tar and robocopy; multilingual injection phrasing.
  • Batch 03–04
    OpenSSL and ssh-keygen; registry exports and queries; AWS and cloud credential stores.
  • Batch 05–06
    USN journal deletion, cipher and WMI persistence; Hermes/ZCode wiring tamper.
  • Batch 07–08
    Split/csplit staging; DPAPI, certificate, autologin and unattended-install files.
  • Batch 09–10
    SSH keys, browser stores, wallets and guarded execution paths.
  • Batch 11
    Terraform, Pulumi, GitHub CLI, Slack, Discord and VS Code state stores.
  • Batch 12
    KeePass .kdbx, 1Password .opvault and .htpasswd outside product folders.
  • Batch 13
    Wildcard path spellings and alternate protected-file patterns.
Keep the OS boundary explicit. The retained 0.61 notes describe a manual firewall block/remove lifecycle. The 0.75 status still reports zero host/OS deny receipts, and the 0.82 note says the kernel/WFP deny path is not enabled. A manual OS-control experiment does not establish integrated Axiom enforcement.

These historical Windows VM results do not establish equivalent enforcement on macOS or Linux, or guarantee protection against unseen threats. No new pentest was run to prepare this page or its documentation.

For the full scope, reproduction checklist and limitations:

Download methodology (Markdown)

Lab screenshot