Verified fixes for the findings your scanner keeps reporting.
AUTOSP, the Autonomous Security Platform, reproduces each finding, patches it in an isolated sandbox, and opens a pull request a person approves.
Under the clouds is the software you run. Each orange crack is a finding nobody has fixed yet.
This is the backlog.
Each crack carries the CWE number of the weakness behind it. The next step lets you pick one and watch it get fixed.
Pick one to fix.
The three choices match the categories the published results cover.
The invoice id goes straight into the SQL string, so a crafted id can rewrite the query.
def find_invoice(cur, invoice_id):
- cur.execute("SELECT * FROM invoices WHERE id = '" + invoice_id + "'")
+ cur.execute("SELECT * FROM invoices WHERE id = %s", (invoice_id,))
return cur.fetchone()
It has to fail first.
Before any fix is written, AUTOSP writes a test that exploits the finding and runs it against your code as it is. If the bug can’t be reproduced, no patch is written.
def test_invoice_id_cannot_change_the_query(db):
row = find_invoice(db.cursor(), "1' OR '1'='1")
assert row is None
$ pytest tests/test_invoices.py -q
F
FAILED tests/test_invoices.py::test_invoice_id_cannot_change_the_query
> assert row is None
E AssertionError: assert (1, 'ACME Ltd', 1200.0) is None
1 failed in 0.21s
The same test later proves the fix.
Written where it can’t reach anything.
A second agent writes the smallest fix it can, inside a throwaway container with no internet, a read-only file system and hard time limits. The container is gone when the job ends.
Drag the sandbox to turn it, or focus it and use the arrow keys.
- Container
- Throwaway
- Network
- None
- File system
- Read-only
- User
- Non-root
- Time limit
- Hard
- Writing the fix in app/repositories/invoices.py
- Patch written: 1 line removed, 1 added
- Not verified yet. The gates come next.
Any one of three checks can send it back.
The reproduction test has to pass now, your existing suite has to pass, and a fresh static scan has to come back clear. The agents can’t reach these checks. A fix that fails one goes back to the sandbox.
- cur.execute("SELECT * FROM invoices WHERE id = '" + invoice_id + "'")
+ cur.execute("SELECT * FROM invoices WHERE id = %s", (invoice_id,))
- Reproduction test passes1 passed
- Existing test suite passes214 passed
- Static scan is clear0 findings
Every merge is made by a person.
AUTOSP opens a pull request with the evidence and stops there. Your reviewers approve and merge under your own branch protection, and the GitHub App can’t change or bypass it.
The careless fix never became a pull request. This is the sandbox’s fix, which passed all three checks.
Press and hold for a moment. Letting go early cancels it.
Merged. You made the call, and the test stays.
Fix SQL injection in find_invoice
autosp/fix-cwe-89 into main
- Reproduction test
- Failed before, passes now
- Existing suite
- 214 passed
- Static scan
- Clear
- Regression test
- Offered in a second pull request
- Review
- 1 approving review required
A record you can’t quietly change.
Every step of your fix was written to an append-only log, and each entry carries the fingerprint of the one before it. Edit any entry and every fingerprint after it stops matching.
Your auditor gets an evidence pack for each assessment.
The fingerprints are real SHA-256, computed in your browser.
Early results, on test applications.
- 14/16
- runs ended with a verified fix (SQL injection, XSS, IDOR and a dependency upgrade).
- 2/2
- misses were closed as inconclusive, with no patch written.
- 0
- merges made by AUTOSP. Every merge is made by a person.
Test applications and a development model, not customer code. A small sample from our 15 October 2026 baseline: an early signal, not a benchmark.
Before you apply
The pilot, week by week
- Week 1Connect
- Week 2First verified pull requests
- Week 4Findings report and evidence pack
To confirm: adjust these weeks to the real plan
What AUTOSP can access
- Your repositories
- Through a GitHub App, on the repositories you choose. To confirm: the exact GitHub App permissions
- Where code runs
- In a throwaway container with no internet, a read-only file system and hard time limits. The container is gone when the job ends. To confirm: where those containers are hosted
- What it can change
- Nothing on its own: it opens pull requests with the evidence and stops there. Your reviewers approve and merge under your own branch protection, and the GitHub App can’t change or bypass it.
- Data retention
- To confirm: how long code and findings are kept
Nothing touches your code until you’ve signed a scope agreement.
How it compares
| What | Scanner alone | Hiring AppSec | AUTOSP |
|---|---|---|---|
| Fixes written | To confirm | To confirm | Yes: the smallest fix it can, in a sealed sandbox |
| Verified before merge | To confirm | To confirm | Yes: the reproduction test, your existing suite and a fresh static scan |
| Human approval | To confirm | To confirm | Always: AUTOSP makes no merges; every merge is made by a person |
| Audit trail | To confirm | To confirm | Yes: an append-only record, and an evidence pack for each assessment |
| Time to start | To confirm | To confirm | To confirm: how long setup takes |
Questions
Which languages does it support?
The pilot is for GitHub-hosted Python services.
Which scanners does it work with?
To confirm: the scanners AUTOSP takes findings from
What can the GitHub App do?
It opens pull requests with the evidence and stops there. Your reviewers approve and merge under your own branch protection, and the GitHub App can’t change or bypass it. To confirm: the exact permissions it asks for
Where does my code run?
Fixes are written inside a throwaway container with no internet, a read-only file system and hard time limits. The container is gone when the job ends. To confirm: where those containers are hosted
What does it cost after the pilot?
The pilot is free. To confirm: the price after the pilot
How soon is the first verified pull request?
To confirm: the usual time to the first verified pull request
Can we host it ourselves?
To confirm: whether self-hosting is offered
For teams with a scanner backlog and no AppSec hire.
Engineering teams of 20 to 150 running a scanner on GitHub-hosted Python services. The pilot is free.
Nothing touches your code until you’ve signed a scope agreement.
To confirm: a two-line note from the founder