AI startups: your API keys are your burn rate
Multi-provider stacks mean multi-provider exposure. One scan covers OpenAI, Anthropic, and every gateway shape.
Your stack, your exposures
AI startups juggle more provider keys than any prior generation: OpenAI projects, Anthropic consoles, gateways, vector DBs — each billable, each tempting to wire straight from the browser during demos.
The three leaks we keep finding here
The demo-direct call
Prototype notebooks graduate to product with keys intact because refactoring felt premature at demo-day.
Gateway migration leftovers
Switching providers leaves old keys bundled “in case we switch back”.
Agent-built features
Assistants calling inference APIs straight from client components.
The check, time-boxed
Set aside ten minutes and run the audit yourself before trusting anyone’s dashboard, ours included:
- Open the deployed site in a browser you do not usually use, logged out.
- View source, then search the built JavaScript for
eyJ— any hit is a JWT; decode its payload and read the role claim before reacting. - Repeat for provider prefixes:
sk_live_,sk-proj-,sk-ant-,AKIA,xoxb-,SG.,whsec_. - Search for connection schemes —
postgres://,postgresql://,mongodb+srv://— and for-----BEGIN PRIVATE KEY-----. - For every hit, classify: public-by-design (anon keys, publishable keys, Firebase web keys) or genuinely secret. Only the second category is an incident.
That routine works, with two caveats: it only covers what the landing page loads unless you chase route-manifest chunks yourself, and it says nothing about yesterday’s deploy. Automation exists precisely because the check decays.
After the first scan
Spend caps + scoped project keys + scheduled scans form the triad; caps bound damage, scoping limits reach, scanning removes surprise.
Boundaries, stated plainly
KeyDrift scans deployed web artifacts — URLs and pasted bundle source. It does not scan mobile binaries, private networks, or repositories; code behind a login is covered by paste mode, and runtime-assembled keys are outside every bundle scanner’s honest reach.
Why this keeps happening industry-wide
Zoom out and the pattern is bigger than one repo. AI-assisted output has outgrown review capacity everywhere at once, which means thousands of teams are making the same reasonable-looking tradeoffs in the same week. Nobody using modern tooling is uniquely exposed. The failure mode documented above is the modal outcome of velocity without verification, not evidence of carelessness.
How KeyDrift reports this exact finding
Report anatomy matters during incidents, so it is worth reading once calmly: masked string (never the live value — it ceases to exist outside the detection engine), salted fingerprint (trackable within your workspace, useless to strangers), chunk path (your starting point for a "git log -S" hunt), disposition (secret versus public-by-design), confidence (matches below 0.5 never reach the page at all).
Manual check, step by step
The full manual drill, for readers who want zero dependence on any tool: open the deployed site in a private window; launch DevTools → Sources; use Search-all-files (Ctrl/Cmd+Shift+F) for eyJ; then repeat for the other marker families — eyJ, sk_live_, sk-proj-, AKIA, postgres, BEGIN PRIVATE KEY. Decode anything JWT-shaped before reacting, and classify public-by-design formats as expected guests rather than intruders.
Close the loop with monitoring
If you take one operational step from this page, make it this: put the URL under continuous monitoring (free tier covers one project daily). The first scan tells you whether you have a problem today; the schedule tells you whether the problem comes back next month after someone re-adds the convenient line.
What the plans change
- Free $0 — 1 project · daily scans · email alerts · findings always visible.
- Indie $29/mo — 3 projects · hourly · Slack added · 80 chunks per scan.
- Team $89/mo — 15 projects · every 15 minutes · Discord + webhooks · 150 chunks.
- Growth — from $249/mo, quoted display-only until checkout ships.
The constant across every tier: plans limit how much is watched, never what a scan found. Visibility is structural, not promotional — asserted by tests over the entitlements model itself.
Common questions
Local model keys too?
If genuinely valueless, severity reflects it — most teams route paid gateways though.
Run a free scan at keydrift.dev/scan — paste a URL or the bundle source itself, no account. Findings arrive masked, with the exact chunk they live in.