Vulnerability Fixer: Automated Dependency Remediation at Jiro Health
- Published on
- • 5 mins read•--- views
Part of my work at Jiro Health is the security audit of the product we build: watching for critical vulnerabilities in third-party libraries, updating them safely, and delivering those patches to every repository. This recording is the presentation of the automated solution I built for that, given to the whole engineering team at the Engineering Round Table on 10 September 2026.
The tool is internal and there are no plans to open-source it, so this post describes what it does rather than how to install it.
The slide from the deck sums up the starting point:
The problem
New CVEs land in dependencies every day. In a HIPAA-regulated product that is not a nuisance, it is an obligation: a known vulnerable package in production is a finding at the next audit. And fixing a CVE by hand is the same ritual every time. Find the advisory, work out which version actually closes it, bump the package without breaking consumers, regenerate the lockfile, open a pull request, update the ticket, tell the team. Then repeat in the next repository.
Three things are wrong with doing that manually:
- Repetitive. Developers spend hours on identical version bumps instead of building the product.
- Does not scale. Every repository has to be checked and patched separately, one at a time.
- Slow. Patching and re-checking everything takes hours, so fixes lag behind disclosures.
What the fixer does
The pipeline runs as a GitHub Actions workflow against each tenant repository: the frontend, the API and the cloud functions. One run does the whole loop.
- Detect. OSV-Scanner and the package manager audit run over every manifest in the repository, including monorepo workspaces, for both npm and pnpm. Results are merged and deduplicated, and the fixed version is taken from the OSV advisory itself, not from the audit tool's often misleading "fix available" flag.
- Respect existing policy. Suppressions and version pins already in the repository are read first, so the scanner never reopens something a human has consciously accepted. It also reports which of those exceptions are no longer needed.
- Remediate in layers. The first layer applies the scanner's own fix. If that is not enough, a second layer forces deterministic pins to the first safe version. A third layer recovers forward to a clean published release and re-scans to prove no new vulnerabilities were introduced. An optional AI write agent can edit manifests in a sandboxed environment with no CI tokens, but it is off by default.
- Verify before committing. Every resolved package is checked against registry integrity digests. A compatibility guard refuses to replace or merge dependency major versions, so a security fix can never smuggle in a breaking upgrade. Anything the guard cannot prove safe goes to manual review instead of into the PR.
- Deliver exactly one PR. The fix branch contains only dependency files. There is always at most one open fix PR per repository: a refresh updates it in place rather than opening a second one, and superseded PRs are closed only after every finding is re-pointed to the new one.
- Close the loop. A Jira issue is created or grown with the findings, moved to In Progress, and linked to the branch. A read-only QA agent writes an impact summary into the PR and ticket body. A Slack summary tells the team what was found and where the PR is.
Findings that have no fix yet are not swept under the rug. A separate suppression-sync run proposes bounded exceptions that expire after seven days by default, in their own PR, for a human to approve. Unknown severities and mixed cases are never suppressed automatically.
Why it matters
The fixer turns vulnerability management from an interrupt into a review. Nobody hunts for advisories, picks versions or rebuilds lockfiles anymore. A developer opens a PR that is already patched, already verified against the registry, already proven not to introduce new findings, with the impact written out, and decides whether to merge.
That changes three things for the company:
- Time. What used to be hours of manual work per repository per disclosure becomes minutes of code review.
- Consistency. Every repository gets the same treatment on the same day, with the same guards, instead of whichever one someone remembered to check.
- Evidence. Every run leaves a trail: findings persisted in a security store, a ticket, a PR with the scan report attached. That is exactly what a HIPAA audit asks for.
Security of the tool itself
A tool that pushes code into every repository is an attractive target, so the pipeline treats itself as untrusted. Runners block egress except to an allowlist. Checkouts never persist credentials. Git operations get short-lived app tokens through in-memory configuration, and those tokens never reach the AI agent or verification steps. The scanner stages only allowlisted paths, never the whole tree, and refuses to publish a PR that touches anything beyond dependency files. Database access uses workload identity with short-lived tokens, with no service account keys and no stored passwords.
The dependency problem is not going away, and this is one of the pieces that lets a small team keep a regulated product patched without turning every CVE into a fire drill.
Related
Projects

Jiro Health, Practice Intelligence for Clinicians
Before: US medical practitioners had to piece together income, referrals and cash flows from systems that were slow to integrate. After: as a principal/founding engineer I own API development end to end for my pods, building enterprise-grade interfaces on Node.js with Python background data processing.
Posts
My Professional Practice Presentation at Teclab
The recorded presentation of my Professional Practice at Teclab, done at Jiro Health on AppSec work, with the slides and the full speech script.
Joining Jiro Health
Since January 2026 I have been working at Jiro Health as a senior backend developer, on a healthtech product for US medical practitioners.

Ekoparty 2026 in Buenos Aires: Three Days of Security Talks in Photos
Photo report from Ekoparty Security Conference 2026: the AI security track, agent and MCP talks, red team tooling, and the installations of the 22nd edition.
Open for contract collaboration
I am available for contract-based collaboration. If you have an interesting project idea, schedule a call via Calendly.
Schedule a 30-min call