Scanner findings in, tested patches out

Point ML.ai at a scanner report. It checks who uses the code, patches each finding, tests the fix, and tells you what only you can do.

Install ML.ai All use cases

New sessionAutomationsReview

Sessions

Upgrade uuid
Review feat/export-csv
Fix scanner findings

Fix security findingsacme/orders9m 21s

Fix the security findings reported in scanner-output.txt

Read scanner-output.txt2 findings

Searched for users of stripeKey and findUserByName0 others

users.js+1−1config.js+1−1

AllowAlwaysDeny

Ran Node with an injection string as the name

The payload goes through as a bound parameter.

One thing I can’t do from here: treat the committed Stripe key as compromised and rotate it.

Describe a task or ask a question

Mode BuildModel AutoEffortlowmediumhigh

src/users.jsPatch

users.jsconfig.js+1 −1

3async function findUserByName(name) {

4- const rows = await db.query("SELECT * FROM users WHERE name = '" + name + "'")

4+ const rows = await db.query('SELECT * FROM users WHERE name = ?', [name])

5 return rows[0]

Who uses for security fixes

Developers with a scanner report and a backlog of small, careful patches.

Teams with a scanner backlog
Findings nobody has had time to patch.
Anyone who committed a secret
A live key that has to leave the code.
Reviewers who want proof
A patch tested against the exploit.

A finding is fixed when the exploit stops working

A patch that reads right can still be wrong.

users.jsconfig.jsscanner-output.txtpasted into chat

3async function findUserByName(name) {

4 const rows = await db.query("SELECT * FROM users WHERE name = '" + name + "'")

5 return rows[0]

The report and the code that calls this never made it into the chat.

What a pasted snippet can see

A chat window sees one pasted snippet

It never sees the full report, the rest of the file, or who else calls the code.

scanner-output.txt:5node-sql-injection [ERROR]

users.js:4db.query("… WHERE name = '" + name + …

config.js:2+stripeKey: process.env.STRIPE_API_KEY

From the finding to the code and its users

It checks who uses the code

ML.ai read the report, opened both files, and searched for callers of stripeKey and findUserByName. None outside those files.

Ran Node with an injection string as the name

SQL: SELECT * FROM users WHERE name = ? ["Robert'); DROP TABLE users;--"]

The payload goes through as a bound parameter, and the key no longer appears in source.

The exploit string, run against the fix

It tests the fix

It ran Node with an injection string as the name. The string went through as a bound parameter.

One action item I can’t do from here: the key was committed to source, so treat it as compromised.

Rotate or revoke it in the Stripe dashboard and set STRIPE_API_KEY in the environment. Removing it from the code doesn’t un-leak it.

What a code change cannot do

It says what a patch cannot fix

A committed key stays leaked. ML.ai told you to rotate it in Stripe and set STRIPE_API_KEY.

Fixing findings by hand versus handing them off

You patch the line the report names and hope nothing else called it. The secret is still in history.

scanner-output.txt, 2 findings

users.js:4 SQL injectionPatched

config.js:2 Stripe keyRemoved

Who else calls these?Not checked

Is the key still live?Open

The ticket closes while the key is still live

You point at the report. ML.ai checks callers, patches, tests, and flags what is left for you.

Fix the security findings in scanner-output.txt

Read the report and searched for callers0 others

users.js+1 −1

3async function findUserByName(name) {

4- const rows = await db.query("SELECT * FROM users WHERE name = '" + name + "'")

4+ const rows = await db.query('SELECT * FROM users WHERE name = ?', [name])

Rotate the Stripe key. I can’t do that from here.

Which agent actually does this

Build

Read reportSearch callersPatchTest

Build · Read report · Search callers · Patch · Test

Patches the findings and tests them

Reads the report and the code, searches for callers, patches each finding, and runs the code to check.

explore

ReadSearchWrite

explore · Read · Search · Write

Optional: see the blast radius

Read-only. Ask where a function or secret is used before anything changes.

From scanner report to tested patch

A finding names a line. The safe fix depends on who else uses it.

Explorer

scanner-output.txt
src/users.js
src/config.js
src/db.js

ML.ai ML.ai

Build

general, the default agent, takes the whole task

Fix the security findings reported in scanner-output.txt

Mode BuildModel AutoEffortlowmediumhigh

1Point it at the reportIn the run, a Semgrep-format text file with two findings. ML.ai reads the report; it does not run the scanner.

Explorer

scanner-output.txt
src/users.js
src/config.js
src/db.js

ML.ai ML.ai

Build

Fix the security findings reported in scanner-output.txt

Read scanner-output.txt2 findings

Read users.js and config.js

Searched for users of stripeKey0 others

Ran Node with an injection string

Describe a task or ask a question

Mode BuildModel AutoEffortlowmediumhigh

2It reads the code and its callersIt opens each flagged file and searches for anything that uses what it will change.

Explorer

scanner-output.txt
src/users.js+1−1
src/config.js+1−1
src/db.js

ML.ai ML.ai

Build

2 findings, 2 patches. The Stripe key still needs rotating, and only you can do that.

users.jsconfig.js+1 −1

3async function findUserByName(name) {

4- const rows = await db.query("SELECT * FROM users WHERE name = '" + name + "'")

4+ const rows = await db.query('SELECT * FROM users WHERE name = ?', [name])

5 return rows[0]

AllowAlwaysDeny

Describe a task or ask a question

Mode BuildModel AutoEffortlowmediumhigh

3It patches each findingIt lists the work, asks which parts to do, and writes each patch as a diff.

Explorer

scanner-output.txt
src/users.js+1−1
src/config.js+1−1
src/db.js

ML.ai ML.ai

Build

2 findings, 2 patches. The Stripe key still needs rotating, and only you can do that.

users.jsconfig.js+1 −1

3async function findUserByName(name) {

4- const rows = await db.query("SELECT * FROM users WHERE name = '" + name + "'")

4+ const rows = await db.query('SELECT * FROM users WHERE name = ?', [name])

5 return rows[0]

AllowAlwaysDeny

Allowed · users.js and config.js patched, tested, key flagged for rotation

Describe a task or ask a question

Mode BuildModel AutoEffortlowmediumhigh

4It tests and flags the restIt runs the exploit string against the fix and tells you to rotate the leaked key.

Explorer

scanner-output.txt
src/users.js
src/config.js
src/db.js

ML.ai ML.ai

Build

general, the default agent, takes the whole task

Fix the security findings reported in scanner-output.txt

Mode BuildModel AutoEffortlowmediumhigh

Where security fixes go wrong

Three common failures, and what ML.ai did about each in a real run.

Callers checkedWith ML.ai

findUserByName0 other users

stripeKey0 other users

Callers checked first

A fix that breaks callers

Changing a function to close a finding breaks code you forgot about.

How ML.ai ML.ai handles it

ML.ai searched for callers of findUserByName and stripeKey first. None outside the two files.

The fix, triedWith ML.ai

users.js:4payload bound as ?

config.js:2no key in source

Tested, not just read

A patch that only looks right

The diff reads like a fix, and nobody tried the exploit.

How ML.ai ML.ai handles it

ML.ai ran an injection string through the patched query. It went through as a bound parameter.

The Stripe keyWith ML.ai

config.jskey removed

Stripe dashboardrotate it

git historyflagged

What is left is named

A leaked secret marked as fixed

The key leaves the code, the ticket closes, and the key still works.

How ML.ai ML.ai handles it

ML.ai said to treat the key as compromised and rotate it in Stripe, which it cannot do from the repo.

Where this stops

What to expect, so nothing surprises you.

Rotate the Stripe keyDo it

Secret rotation is yours

ML.ai changes the code and tells you to rotate the key. It cannot reach Stripe.

Also changing a file it did not declare

AllowAlwaysDeny

It stops at undeclared files

Editing a file outside its stated list needs your OK.

exploreBuild

Check the blast radius first

Use @explore, which is read-only, to see where a function or secret is used before patching.

uuid 3.4.0 to 14.0.3

npm install uuid@latest14.0.3
require("uuid") in this CJS projectv4 works
npm test2 passed

ids.js+1 −1

1-const uuidv4 = require('uuid/v4')

1+const { v4: uuidv4 } = require('uuid')

Upgrade dependencies

ML.ai bumps a package, fixes the code the new version breaks, and runs your tests.

See how it works

Root cause in session.ts

The store returns null before the session write finishes, so the middleware rejects a valid token.

session.ts+2 −1

40export async function getSession(id) {

41- const s = store.get(id)

41+ const s = await store.get(id)

42+ if (!s) throw new SessionNotFound(id)

AllowAlwaysDeny

Debug

Give ML.ai a failing test or a stack trace. It finds the actual cause and proposes a fix you approve.

See how it works

FAQs

What does ML.ai do with scanner output?

It reads the report you point it at and works each finding: open the code, check callers, patch, test. It does not run the scanner.

Does it check that the fix works?

Yes. It ran the patched query with an injection string and showed it bound as a parameter.

What about a committed secret?

It moves the secret out of the code and tells you to rotate it. Removing it from source does not un-leak it.

Where does my code go?

Per ML.ai’s docs, prompts, selected file content and tool results go to the ML.ai model service; the engine runs locally. Check your organization’s policies for confidential code.

Try today, or talk to us about what is next

Install the editor agent on your own machine, or book a call to talk through your team’s workloads

Download ML.ai Book a call

Scanner findings in, tested patches out

Who uses for security fixes

Teams with a scanner backlog

Anyone who committed a secret

Reviewers who want proof

A finding is fixed when the exploit stops working

A chat window sees one pasted snippet

It checks who uses the code

It tests the fix

It says what a patch cannot fix

Fixing findings by hand versus handing them off

By hand

With

Which agent actually does this

Patches the findings and tests them

Optional: see the blast radius

From scanner report to tested patch

Where security fixes go wrong

A fix that breaks callers

A patch that only looks right

A leaked secret marked as fixed

Where this stops

Secret rotation is yours

It stops at undeclared files

Check the blast radius first

Related use cases

Upgrade dependencies

Debug

FAQs

Try today, or talk to us about what is next