Finding vulnerabilities is easy. Closing them is the work.
AUTOSP takes one finding from report to a pull request with proof. A person decides what merges.
-
Find
It scans the repositories you choose, or takes results from the scanner you already run.
- Semgrep
- CodeQL
- Snyk Code
- osv-scanner
CWE-89
app/repositories/invoices.py -
Reproduce
An agent writes a test that exploits the finding. The test has to fail. No failing test, no patch.
# the exploit, as a test def test_find_invoice_injection(client): r = client.get(INJECTION) assert r.status_code == 4041 failed in 0.21s
-
Fix
A second agent writes the smallest fix it can, in a sealed copy of your code.
- No network
- Read-only
- Non-root
- "… WHERE id = '" + invoice_id + "'" + "… WHERE id = %s", (invoice_id,)
-
Verify
Three independent gates check the fix. The agents can't see or change them.
- The exploit test now passes
- The full test suite passes
- A fresh scan finds nothing new
-
Review
AUTOSP opens a pull request with the evidence and stops. Your reviewers merge it.
Pull request #412Waiting for you
Fix SQL injection in find_invoice
MergeMerges made by autosp: 0
Three gates. Any one of them can send a fix back.
-
The exploit fails now
The test that failed before must pass on the patched code.
before: failednow: passed
-
So do the variations
Attack variants run against the fix, judged by rules AUTOSP owns, not by the agent that wrote it. For SQL injection: 18 variants.
-
Nothing else broke
Your test suite still passes, a fresh scan no longer reports the finding, and your tests and CI are untouched.
- Test suite
- Fresh scan
- Tests and CI untouched
If a fix can't be proven, AUTOSP says so. It's never called fixed without evidence.
AUTOSP never merges. Not even when every check passes.
- entry 41
sha-256 9c1e…07ab - entry 42
previous 9c1e…07ab - entry 43
previous 4d82…f31c
-
Only the repositories you pick
A GitHub App with four permissions: Contents, Pull requests, Checks and Metadata. It never asks for Administration, so it can't touch your branch protection.
- Contents
- Pull requests
- Checks
- Metadata
-
A sealed sandbox
Every job runs with no network and a read-only file system, on a disposable copy that is deleted afterwards.
- No network
- Read-only
- Non-root
- Time limit
-
Little goes to the AI model
It sees the finding and the specific files it needs, never your whole repository or your Git history.
-
Every step on the record
An append-only log where each entry carries the fingerprint of the one before it, so a change can't go unnoticed.
Early results, on our test applications.
| Run set | Date | Verified |
|---|---|---|
| SQL injection | 22 Oct 2026 | 5 of 5 |
| Access control (IDOR) | 22 Oct 2026 | 5 of 5 |
| Mixed run set | 2 Nov 2026 | 15 of 15 |
0 merges made by AUTOSP, and known-bad fixes rejected every time.
Read this before the numbers. These are test applications and a free development model, not customer code and not the production model. Small samples: an early signal, not a benchmark.
Try it on your own code.
For Python teams on GitHub, of 20 to 150 engineers, with no dedicated security engineer.
- You sign a scope document naming the repositories
- Your reviewers handle AUTOSP's pull requests
- Your tests run offline
Nothing touches your code until you've signed a scope agreement.
Questions teams ask.
Does AUTOSP replace my scanner?
No. Its job is closing findings, not finding more of them. Keep your scanner; AUTOSP fixes what it reports.
Can it merge by itself?
No. Not in any configuration. Your reviewers merge.
What does it cover today?
Python on GitHub: SQL injection, cross-site scripting, insecure direct object references and vulnerable dependencies. It works with Semgrep, CodeQL, Snyk Code and osv-scanner.
What if it can't fix something?
It says so: inconclusive, needs human review, or a recommendation clearly marked as advice.
Will a passing scan mean my app is secure?
No. "No findings from these scanners" is all a scan can say, and that's how AUTOSP reports it.