A website can deliver an authentic script from its usual supplier and still expose visitors to malicious code. That is the central failure in the Optimizely Web Experimentation incident of 16 November: trust in the delivery address outlasted trust in the content being delivered.
Optimizely's 17 November initial incident report says a signup-flow flaw allowed an attacker to impersonate users and modify customer Project JavaScript. It says customer passwords were not stolen or used. The affected surface was Web Experimentation, rather than Feature Experimentation, SDKs or legacy Full Stack.
The delivery address stayed legitimate
In its original investigation published on 16 November, Report URI describes malicious code inserted into customers' Project JavaScript. The surrounding snippet still came from Optimizely's trusted CDN. The researcher explicitly distinguishes this from CDN cache poisoning.
That distinction changes the incident model. A delivery network can correctly serve the object it received while the system authorized to change that object has been abused. Availability monitoring, a valid certificate and an unchanged script address do not establish that the script's behavior is still acceptable.
KrebsOnSecurity's independent reporting on 17 November includes an Optimizely statement confirming remediation and an ongoing investigation. The incident is therefore supported by both the supplier's account and an external investigation, rather than a claim inferred from a website outage.
Infected snippets are not a count of compromised visitors
Report URI found 188 infected snippets associated with 178 organizations among 1,611 snippets it inspected. Those are observations of distributed code, not a census of people who executed it or lost data.
Its analysis describes payment-card collection, cryptocurrency-address substitution and other browser payloads. A visitor's actual exposure depended on the page, device, activity and browser policy. An affected organization's name alone does not establish that every visitor was compromised.
This distinction is essential for incident reporting. Evidence that a supplier distributed a malicious object answers a distribution question. Establishing individual impact requires additional evidence about execution and subsequent activity. Multiplying an organization's audience by an infected-snippet count would produce a speculative victim total.
CSP can stop a stage without neutralizing the whole attack
Content Security Policy controls several separate browser capabilities. In the W3C's CSP Level 3 working draft, script loading, connection destinations and frame loading have separate directives. A report-only policy monitors violations without enforcing the restrictions.
Consequently, saying a site “has CSP” is incomplete. A policy may allow the initial supplier script while blocking a later payload from another origin. It may permit code to run while restricting its outbound connections. Those are different outcomes with different residual risks.
Report URI observed a partial outcome in which outbound data collection was blocked but cryptocurrency-address substitution still worked. That is a particularly important boundary: changing information already displayed in a page can cause harm without sending stolen fields to a collection server.
CSP's strict-dynamic mechanism also deliberately permits dependencies loaded by appropriately trusted scripts. It can simplify dependency management, but that trust propagation matters if an admitted script becomes hostile. Neither the presence of that keyword nor a long domain allowlist establishes protection against this particular incident.
The evaluation has to follow the actions the malicious script attempts. Whether the browser admitted the first script, admitted its next dependency and allowed an outbound request are separate pieces of evidence. A blocked request is valuable evidence of a blocked action, rather than proof that every harmful action was impossible.
Integrity checks answer a different question
Subresource Integrity lets a browser compare a fetched resource with expected cryptographic integrity metadata. This checks the resource's bytes against an expected value, rather than relying only on its location. The current W3C document is a working draft; it should not be read as a claim that every browser implements every proposed feature.
For a deliberately changing experiment snippet, keeping an expected digest stable creates an operational tradeoff. A legitimate content update changes the bytes too. An integrity mismatch can prevent use of the changed resource, but the organization still needs a trustworthy decision about which new version to approve.
Automatically accepting every new digest supplied through the same compromised change path would preserve the appearance of verification while losing the independent approval decision. That is a design implication, not evidence that Optimizely offered or bypassed a specific integrity implementation.
Remediation and the remaining evidence
Optimizely's initial report says it blocked malicious outbound code at 16:06 UTC, fixed the flaw at 16:10 and removed injected code from all affected projects by 22:24 on 16 November. These are the vendor's milestones, not a universal per-visitor exposure window. Its report distinguishes card entry in the site's own forms from hosted payment iframes that the script could not read.
Removal of the malicious code ends a delivery path; it does not undo previous execution or prove the absence of loss. A complete account must separate supplier remediation from findings about visitors, payments and sessions.
Our NEAR AI incident analysis examines a different trusted-software boundary inside confidential inference. Here, the browser admitted a dependency through the website's own supplier relationship. The consequential decision was which remotely changeable code could act inside the page, and which independent controls still constrained it after that trust failed.

