Code security
in the AI era.
How many of you used AI
to write code this week?
Keep your hand up if you reviewed it as carefully as code you wrote yourself.
"vibe coder"
over
year
Not because researchers got better at finding bugs. The attack surface is expanding faster than the defenses.
Vunerability Trends
tj-actions/changed-files
A single GitHub Action thousands of CI pipelines already trusted. The attacker didn't need your repo — just the tag you were already pulling at runtime.
Trust was the delivery mechanism.
Your security tool became the attack vector.
Axios
A JavaScript HTTP library running in 80% of cloud and code environments.
North Korean threat actor compromised a maintainer's npm account. Malicious versions lived for hours — long enough at 100M weekly downloads. Payload: a backdoor called Waveshaper.v2 across Windows, Linux, and Mac.
/ week
Mini Shai-Hulud · TanStack, UiPath, MistralAI
versions
Rotating your credentials triggers the bomb.
You're not just
responsible for the code
you write.
You're responsible for the code
you trust.
Move security
to the moment of writing.
Not the moment of review. Not the moment of CI. The moment the code is typed.
Pin actions to a commit SHA.
Mutable tags are the root cause of tj-actions. The roadmap locks every action — direct & transitive — like a go.sum for CI, plus runner-level egress firewall and org-wide execution policies.
Three to enable now.
Free for open source. On GHAS for private repos.
Dependabot
CodeQL
Secret scanning
Catch it before the commit.
Secret scanning is GA via the GitHub MCP Server. Dependency scanning is in public preview. Your agent scans while you type — not in CI, after the fact.
/secret-scanning scan my current changes for exposed secrets
Fix in the PR. Triage by reality.
Copilot Autofix
CodeQL flags a vulnerability. Autofix shows you the fix inline. PR suggestion you accept, edit, or reject.
Defender for Cloud
code-to-cloud · GA
Filter CodeQL alerts by what's actually running and reachable.
The data isn't encouraging in the abstract. The tooling is real.
- —CVEs are up.
- —Supply chain attacks are up.
- —TeamPCP campaigns are still active. The TanStack blast radius is still expanding.
- +Scanning in the agent, before the commit.
- +Fix-suggestion inside the PR, not a JIRA ticket.
- +Runtime context: what's actually deployed.
Are we ahead or behind?
- •More vulnerabilities are being surfaced than ever before.
- •Better tooling is improving early detection and response.
- •Volume, triage pressure, and low-quality reports still create drag.
- •Strong security practice must be practical, clear, and accessible.