Stroq shows when a coding AI is tricked into leaking keys
Stroq traces risky coding AI commands back to outside text. In my fake session, it would block three of five actions, including sending a fake key away.
Head of Product at Sprouts.ai
A coding assistant may read a bug report or web page while fixing your app. If that page hides an instruction to send away your passwords, the assistant may follow it. Stroq, from AGGIB, checks the assistant's saved history and can check commands before they run.
Stroq checks commands before they run in Claude Code, Cursor, Codex, Copilot CLI, Windsurf and Antigravity.
I planted a malicious bug report and a fake API key, a secret used to access a service, in a saved Claude Code history. Stroq traced the path from a request to fix checkout to commands that ran downloaded software and sent the secret elsewhere. Before giving a coding tool real keys, a PM can ask the team to show which outside text caused each risky action.
This was a fake session with a fake key, so these are what Stroq would block, not things it blocked live.
What I tested
I planted a checkout bug report telling the assistant to run a script from evil.example and send its secrets file to collector.example. The saved history included a fake Stripe key, a command sending that key away and a final test command. I asked Stroq to find the suspicious instructions, trace the commands back to their source and check for the exposed key. The saved package identifies itself as v0.31.1.

Stroq's summary card points to the script command and names the fake key's destination.
What would Stroq have stopped?
I used replay, which means rerunning the saved history through Stroq's rules without executing anything. It judged five actions and would block three:
Reading the secrets file after encountering the suspicious issue.
Downloading a script and running it immediately.
Sending the fake key to the address in the issue.
It would allow opening the issue and running the final test command. The detailed view connected both risky commands to the issue text and showed where the fake key came from. I could see the instruction that caused the problem alongside the action it led to.
The card, secret check and replay summary each finished within about a second. Stroq's separate set of 36 built-in attack examples took about two seconds.

The detailed view names the rules: deny-secrets-when-tainted for reading the secrets file, deny-encoded-exec for downloading and running the script, and deny-secret-egress for sending the fake key away.
How would I start?
Stroq is free and open source under the Apache-2.0 license. It runs locally and needs no account. For a PM, I would start by asking an engineer to run it on one recent session and walk me through the card.
The official instructions require Node.js 22 or newer, the software that runs this tool. Install that first. Then open a terminal in your project and run:
npx @stroq/clinpx fetches and runs the tool. This first command checks past sessions and prints the summary card. That card is a good first result. To add checks before future commands, run:
npx @stroq/cli initThat command changes the coding assistant's settings. Stroq calls these checks “hooks”: checks that run before the assistant's next command. Adding them to settings does not prove the assistant calls them. To remove Stroq, run:
npx @stroq/cli uninstallOpen Stroq's local control page with:
npx @stroq/cli consoleUse its “Test hook” control; that is the sign the checks are really connected. The docs say this runs the check once, offline. The “Live” view shows decisions as they happen. I would ask the engineer to show those results before relying on the setup.
How did a bug report become a security problem?
Prompt injection means hiding instructions in material the assistant is meant to read as information. A bug report can explain a checkout failure. In my example, it also told the assistant to download software and send a secrets file to an outside address.
The file was named .env, a common place to store app settings and secrets. The issue requested the whole file; the recorded outgoing command contained the fake key from it. Stroq connected that command to both the issue and the file read.
I would want that explanation in a review. “Fix checkout” sounds like normal development work. The instruction to send a key to the issue author's chosen address is something the team should reject, even if the author presents it as a debugging step.

The five actions in order, with Stroq's decisions beside them.
What did the built-in attacks show?
I also ran stroq attack. It checked the default rules against 36 built-in examples: 26 drawn from public reports and ten made-up examples. Some adapt a reported incident rather than reproduce it exactly.
Stroq reported 15 blocked, 21 asked about and zero passing through. More than half the risky actions stopped to ask a person instead of being blocked. Someone has to answer those prompts, and they need to see the outside text next to the command to answer well.
These results show how Stroq handles its bundled examples. They do not establish how often it will catch every new attack. I would use the examples to review which decisions the rules make and which ones the team is leaving to a person.

Twenty-one examples required a human answer. That is review work the team needs to plan for.
What should I be careful about?
The secret check named STRIPE_SECRET_KEY, its source file and collector.example as the destination without printing the key's value. To report a definite send, it looks for the key in an outgoing command. A script or file can carry a key away without showing its value there, so a clean report does not prove nothing leaked.
I also found a counting mismatch. The card said one command came from text the assistant read. The detailed history view traced two: the script command and the fake-key upload. The saved output does not explain why the counts differ. I would inspect the detailed view rather than rely on the card's total.
Stroq checks proposed actions against rules; it does not put running code inside an isolated environment that limits what it can reach. That is what “a filter, not a sandbox” means here. It also depends on the coding tool passing actions to its checks, and it traces source text by matching words and values. The integration limits are worth reviewing with the engineer responsible for setup.
For the separate question of isolating running code, my OpenShell breakdown covers that approach.
What this means for product and GTM teams
I would make the issue-to-command explanation part of the review before a coding tool gets real credentials. Ask the team to show one recent session, open each risky action and point to the outside text that led to it.
That gives a PM something specific to question. Why did a checkout fix need to send a secret to an unfamiliar address? Who would answer the approval prompt? Does that person see the original instruction or only a command?
The approval screen deserves the same attention as the scanner. My agent-autoloop test covered a different approval workflow. Here, I would ask to see the issue text beside the command, then have the intended reviewer explain why they would approve or reject it.
Frequently asked questions
Did Stroq block a real leak in this test?
No. I checked a fake session with a fake key. Stroq said it would block three of five actions; the check did not execute those actions or send the key anywhere.
Does zero passing through mean every attack was blocked?
No. Fifteen built-in examples were blocked and 21 required a person to answer. Those questions still need review.
Does Stroq show the secret value in its report?
My report printed the key's name and destination, not its value. The official docs say it prints names rather than values.
Can a PM try it without learning the commands?
I would ask an engineer to run the first check and walk through the card. The useful questions are which outside text led to each risky action and who would approve it.
Sources
Stroq repository and README: starting commands, no-account setup, checks before commands and limits. Checked 10 November 2026, IST.
Package details: v0.31.1, Apache-2.0 and Node.js 22 or newer.
My saved card, secret check, attack results and detailed history view, shown and described above.

Kartikay Dhar
Head of Product at Sprouts.ai, where he leads One Platform, Sprouts Data Universe and the Sprouts MCP. On Instagram he shares one free GitHub repo a day.


