Skip to evidence

INDEPENDENT AI CLI NETWORK TESTS · Aug 14, 2026

What did your AI coding tool contact before the work even began?

We rerun clean, fixed scenarios against exact CLI builds. Every result states the version, config, repeat count and limits. No privacy score. No guess about intent.

CHECK YOUR TOOL READ THE CLAIM CONTRACT
THE ANSWER IN ONE LINE

A result describes only what this exact build did in this exact test. It does not establish purpose, retention, training use, sale, safety, or behavior in another account, region, operating system, or scenario.

CURRENT COVERAGE

3 exact builds. 12 primary runs.

RETEST QUEUED

Anthropic · 2.1.232

Claude Code

The documented opt-out removed every observed nonessential route in 3/3 fixed-window runs.

Upstream 2.1.283 detected. Published evidence still describes 2.1.232.

REPEATS
3/3
DECODED HOSTS
2
REQUESTS / RUN
9–29
INSPECT
RETEST QUEUED

OpenAI · 0.147.0

Codex CLI

Startup contacted OpenAI and GitHub infrastructure in 3/3 final runs; two OpenAI routes varied across batches.

Upstream 0.157.1 detected. Published evidence still describes 0.147.0.

REPEATS
3/3
DECODED HOSTS
5
REQUESTS / RUN
20
INSPECT
UPSTREAM CHECK MANUAL

Cursor · 2026.03.30-a5d3e17

Cursor CLI

Invalid-key startup reached one Cursor host through two routes in 3/3 runs.

REPEATS
3/3
DECODED HOSTS
1
REQUESTS / RUN
2
INSPECT

WHY RETURN

Each release can change the answer.

01New CLI build detected.

02Same isolated scenarios rerun three times.

03Version diff and egress manifest published.

04The public version diff and machine data update together.

ONE FOUNDING PLACEMENT · MEASUREMENT CANNOT BE BOUGHT

Security product in front of developers deciding what may leave their machines.

One contextual placement outside evidence tables. Public impressions and clicks. Sponsor receives no retest, ranking or editorial control.

SEE THE CONTRACT