Audit Third-Party Scripts Before They Reach Candidate Data

Run a third-party script audit across application forms and candidate portals. Test with dummy data, review browser controls, and describe exposure precisely.

Ernest Bursa

Ernest Bursa

Founder · · 9 min read
Two colleagues on a San Francisco rooftop discussing a paper map beside a closed laptop; one wears a navy Startupkit crewneck.

A third-party script audit should follow candidates from the job description through the application form and into their portal. List what runs on each page, who can change it, and why it needs to be there. Then test with synthetic data and distinguish what a script could access from what you observed it sending.

That distinction matters when a trusted browser dependency is implicated in an incident. A vendor name in a script inventory tells you where to start looking. It does not tell you which pages ran an affected version, which payload executed, or whether candidate information left the browser.

The useful question for a hiring team is concrete: which outside publishing systems can put executable code beside our candidates’ information?

What the Optimizely incident does and does not establish

On 16 November 2026, Report URI reported finding 188 infected Optimizely snippets for 178 organisations among 1,611 snippets fetched. Those are results from its discovered sample, not an estimate of all customers. Its investigation described injected Project JavaScript in trusted snippets, including customer-hosted copies, and said the initial write-access method remained unknown.

Amirul Iman’s independent investigation also documented malicious payloads and instances where CSP blocked attempts. Neither investigation establishes theft of candidate records or involvement of Kit. Report URI’s late 16 November observation that most checked snippets had been cleaned is a dated finding, not assurance about every deployment today.

The hiring implication comes from the browser’s trust model. As OWASP explains, third-party JavaScript executing in a page can access its document. The label “experimentation” or “analytics” does not confine that code to experiment results or approved analytics fields.

Follow the candidate journey, not just the careers homepage

Audit each place where the data or authority changes. A careers homepage, an application form, and a candidate portal may share a layout while exposing very different information.

Start with a page inventory. Include redirects, error states and pages reached through email links. Check the signed-in recruiter workspace separately: a public-page inventory cannot establish what runs beside internal hiring records.

Journey surface Data or authority to inspect Decision to make Verification
Careers listing and job description Public content, URL parameters and any personalised content Does this script need to run on this route? Capture loaded scripts and their initiators on the actual page.
Application form and upload Typed contact details, answers and selected files Which scripts are necessary to complete the application? Exercise typing, errors, file selection, upload and submit with dummy data.
Receipt and candidate access Access links, redirects and application information Can optional code be removed from these pages? Follow the complete access journey and inspect URLs and requests.
Recruiter workspace Records visible to the user and actions their session permits Does the vendor need this level of browser access? Repeat the audit in a test account with representative permissions.

Do not assume a page is low risk because it sits under a careers URL. Inspect its contents. Equally, do not claim a script can read arbitrary files on a candidate’s computer. Review the files selected or made available through the actual application flow.

Give every script an owner and a reason to run

For each dependency, record a business owner, a technical removal owner and the people or systems able to publish changes. A script URL alone leaves the most important question unanswered: who controls the next version?

Include scripts inserted by tag managers, experiments and other scripts. Record their initiator chain: the sequence that caused the browser to load them. A short HTML source file can still trigger a long chain of scripts after the page opens.

Ask the owner to name the specific function the script performs on that route. “We use this vendor” is too broad. An experiment on job-description wording does not establish a need to execute the same dependency inside candidate access pages.

Removing a script with no necessary function reduces exposure. Where it is necessary, document its publishing path, controls and removal procedure. Test whether turning it off breaks an application, upload or access link before relying on that procedure during an incident.

Self-hosting also needs a precise description. A reviewed, fixed copy under your release process differs from a proxy that keeps fetching mutable vendor content. Your domain in the address bar does not establish that the code has been reviewed.

Test access and network traffic with synthetic data

Use a staging account and recognisable dummy values to observe the journey. This is a proposed audit method, not a report of tests performed on your deployment.

  1. Create a dummy candidate. Use distinct marker values for name, email and answers, plus a harmless test document. Keep real candidate records out of the test.
  2. Record the conditions. Note the time, route, application build, authentication state, consent state and script URLs or available versions. Different conditions can produce different behaviour.
  3. Capture the loading chain. In browser developer tools, inspect network requests and their initiators. Follow code loaded after the initial document, including after an interaction.
  4. Exercise the whole flow. Type into fields, trigger validation errors, select and upload the dummy file, submit, inspect the receipt, and follow candidate access. Check the states your candidates actually encounter.
  5. Inspect destinations and contents. Where visible, check request URLs and bodies for your markers. Record a matching value alongside the destination, time and initiator. A request to an unfamiliar host is a lead to investigate; it is not by itself proof that candidate data was sent.
  6. Repeat after a control change. Confirm the intended request is blocked or the dependency is absent, then verify that legitimate forms, uploads and access still work.

Treat capture files as sensitive. A browser network archive can contain cookies, access links and request bodies. Do not publish raw captures or upload real records to an outside scanner to make the audit easier.

A clean run has a useful but bounded meaning: you did not observe the tested transfer under the recorded conditions. Conditional payloads, different routes or a later remotely published version may behave differently. Keep that limit beside the result.

Know what CSP and integrity checks can actually stop

Content Security Policy, or CSP, can restrict browser resource loading and execution. Its effect depends on the policy the page receives and whether the browser enforces it.

Report-only CSP does not block a violation. OWASP’s CSP guidance distinguishes observation from enforcement. Check the response headers and the actual browser result before writing “CSP prevented execution” in an incident note.

An allowed origin is also a trust decision. If code from that origin changes, an origin allowlist does not review the new code. A policy using strict-dynamic can extend trust from an authorised script to scripts it loads; unsafe-eval permits text-to-code execution. Assess these settings against your page’s purpose and dependencies.

Optimizely’s own CSP documentation describes nonce and strict-dynamic compatibility, with unsafe-eval needed for some custom-JavaScript conditions. Compatibility instructions explain how to support features. You still need to decide whether those features belong on a page holding candidate information.

Subresource Integrity, or SRI, answers a narrower question: do these fetched bytes match the expected hash? MDN documents that a mismatch prevents execution, and that cross-origin use requires CORS support. A changing experiment bundle therefore needs a deliberate review and hash-update process. Pinning one loader does not automatically pin everything it later fetches.

Nor should masking or an approved analytics data layer be described as a sandbox for arbitrary JavaScript. Separate-origin frames can provide a different boundary, but their permissions and messaging need their own review. These controls solve different problems; verify the specific promise you intend to make.

Bound the exposure before you describe it

Record dependency, execution and transfer findings separately. Each answers a different question and requires different evidence.

Finding What supports it What it does not establish
An affected dependency was deployed A dated script version or captured bytes tied to a route That every visitor received or executed the payload
Code executed beside candidate information Browser evidence connecting execution to a page and state That the code transmitted that information
A transfer was attempted A captured attempt, including any observed browser block Successful delivery or receipt
Candidate-like values appeared in an outbound request A captured request carrying the synthetic markers The extent of historical real-candidate exposure

For example, “the application page loaded an affected script during this release window” is a meaningful finding. “All applicants’ résumés were stolen” needs substantially more evidence. Missing historical logs cannot close that gap in either direction.

Preserve the route, time window, relevant versions, observed destinations and control behaviour. When you remove a dependency or block a destination, record what changed. Blocking a request does not undo changes code has already made inside the page.

Keep the finding connected to the hiring workflow

Reviewing browser dependencies is part of protecting candidate information. Kit provides careers pages, application forms and candidate access, so those surfaces need the same explicit audit decisions. Its Security report workflow can help record and follow up a finding; it does not replace the browser investigation or establish containment.

If you are reviewing the wider data journey, use Candidate Privacy & Consent and the guide to securing candidate PII to connect the script review with the information your team collects and retains.

Start with one complete application journey. Give each dependency an owner, remove what has no necessary role, and retain enough test evidence to explain what you observed. That gives your team a concrete decision today and a defensible starting point when the next vendor alert arrives.

Related articles

Try Kit for 30 days.

Hiring, security reports, and training in one account, for teams where none of it is a full-time job. Free for 30 days, card required. Cancel before it ends and you pay nothing.

Get started free