Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Experiment result

Stroq attack suite

Written November 8, 2026

We ran Stroq's own attack command to see whether its hooks stop the attacks it ships with.

What was tested

Stroq guards coding agents with local hooks that judge each tool call before it runs. Its attack command runs a set of known attacks against your policy in a throwaway home directory, to show what gets stopped and what gets through. We tested whether the command's result on version 0.30.0 matches what the README says it should be.

How

One command, on version 0.30.0 of the Stroq CLI:

npx -y @stroq/cli attack

The attacks run in a sandbox the command sets up itself. Nothing else was installed and no agent session was involved.

What happened

OutcomeScenarios
Blocked (denied)15
Asked (stopped for a human decision)21
Passed through0
Total36

That matches the README's claim for 0.30.0 exactly: "on 0.30.0, 15 are denied, 21 are asked about, 0 pass through." The README also says the 36 attacks come from public reports and synthetic test cases.

What it means for the reader

Stroq's own self-test passes on its own terms: nothing in its attack set got through unchallenged on this version. But most of the set — 21 of 36 — was not blocked outright; it was stopped and handed to a person to decide. The protection there depends on the person saying no, so the result is a prompt to check your own habits, not a promise that nothing can slip past.

Limits

This is a self-test, not an independent evaluation. The 36 attacks are Stroq's own set, so the run shows Stroq stopping attacks Stroq chose, not attacks an outsider devised. No transcript audit was run: there were no real agent sessions on this machine, so the result says nothing about what Stroq would catch in a live session, or about attacks outside its set. A new version may change the counts.

Related experiments

Sources

Back to: Stroq: audit what your coding agent touched