AdFection: the gap between an approved ad and the code a device runs

1. From approval of an ad to the code actually delivered

Citizen Lab’s October 15, 2026 AdFection report links Al Jazeera infections to Candiru through Discover-Tech ads, using network sequences and a device sample. Its 45-plus targeted devices are not a confirmed infection count. [1]

The question for advertising markets is what a review covers. Appropriate imagery and wording do not describe the entire user experience if a service called by the ad subsequently supplies different code. An ad slot is sold as part of a screen, but the relevant quality extends beyond what appears on that screen. The commercial unit and the unit within which technical effects occur can diverge.

A service that approves a submitted creative once and one that verifies subsequent changes offer different assurances. If buyers regard them as equally safe, providers funding broader verification may be disadvantaged on price. If the tested conditions and scope are visible, continuing verification becomes a quality buyers can compare. SG Group reads this difference in assurance as a competitive condition in advertising intermediation.

The initial effects concern ad delivery and the protection of browsing devices. The reports cannot yield a percentage decline in global advertising expenditure. Revenue and spending implications become concrete when customers revise verification scope, inventory selection or purchasing terms. Following who verifies quality and who can use the result reveals differences between companies more clearly than jumping from the existence of a threat to aggregate industry profits.

2. What counts as the same advertisement?

Advertisements can be grouped by image, identifier or responding provider. Consider an illustrative service that keeps the initial image unchanged while tailoring its subsequent response to recipient conditions. One creative identifier could then encompass different behaviour. Retaining the submitted creative remains useful, but it does not freeze later responses.

Variation alone cannot establish misconduct. Adapting presentation to display size or device capabilities can have legitimate purposes. Verification must define the permitted range of variation and detect behaviour beyond it. Treating a presentation adjustment and an additional outward transfer of information as equivalent updates obscures meaningful differences in quality.

A benefit of distributing one creative widely is efficient updating and optimization. Preserving that benefit requires clarity about which changes remain within approval and which require renewed verification. An illustrative contract might distinguish resizing an image from adding an external service. A rule for continued use of an approval identifier allows review effort to follow consequential changes rather than every minor edit.

The guide to advertising claims and approval records explains how to record who reviewed which version. Substantiated wording and acceptable runtime behaviour are distinct outcomes. Buyers needing both must specify the corresponding deliverables and update conditions. A process in which editorial approval is mistaken for completion of technical verification leaves a gap in responsibility.

3. Responses a review device may never receive

Confiant’s investigation, published the same day, examined a framework that profiles devices and selects subsequent code. Its observation window ran from April 1 to September 10, 2026. An available capability does not establish its use on every device; the execution context also constrains it. An impression on an organization’s network does not identify a targeted or infected employee. [2]

For a verifier, differences between responses under distinct conditions matter more than one successful test. A clean review device leaves open whether another recipient receives the same response. If the sender can distinguish recipient conditions, merely repeating the review does little to test that assumption. More observations need not cover a broader range.

Compare one hundred loads on one review device with ten loads across each of ten distinct conditions. Both total one hundred, but the second design addresses variation between conditions. Repeated tests under one condition can still reveal changes over time. The purpose determines the allocation: investigating device variation and investigating temporal variation require different designs.

This comparison estimates no detection probability. Ten chosen conditions may not represent actual recipients, and a sender may recognize review environments. It nevertheless shows why selling a raw test count can favour inexpensive repetition within one narrow range. Asking about condition selection allows verification providers to compete on reducing blind spots rather than maximizing repeated observations.

4. Why impression counts cannot become victim counts

Figure 3 compares different transaction stages, not people or infected devices. Wins divided by bids equal about 11.7%, not an infection rate. Large denominators require care about the conditions governing movement to the next stage. [1]

Commercial loss and conveniently measured delivery counts are different quantities. In an illustrative case, one person may receive an ad repeatedly, raising impressions without increasing recipients. A shared device creates a further gap between device and person counts. Delivery records cannot transfer unchanged into estimates of compensation or the number of devices requiring investigation.

Basing budgets only on confirmed infections creates another problem: weak investigative capacity can look like low harm. A thoroughly examined organization may appear less safe than one lacking records simply because visibility differs. A target to reduce detections can even be met by investigating less. Evaluating defensive investment requires avoiding that incentive.

Delivery, response, execution and device impact require different evidence. Earlier-stage records can prioritize later investigation without establishing its outcome. Buyers need to know which stage a service documents. Comparing delivery logs with device investigations as if they were equivalent products can leave the required capability unpurchased despite the reassurance of having a contract.

5. A legitimate selling path and acceptable execution

IAB Tech Lab describes ads.txt as a publisher’s declaration of authorized inventory sellers. sellers.json and the SupplyChain object add visibility into sellers and resellers. The respective pages display updates of September 14 and October 7, 2026. Their stated purposes concern selling authority and the supply path rather than verification of every behaviour of delivered code. [3] [4]

Identity checks remain valuable. Discovering an abnormal response along an unknown selling path can leave no clear recipient for a suspension request. Identification enables investigation, contact and contractual action. It should not be marketed as proof that every response was safe. Combining mechanisms with different functions expands what can be established.

A publisher may know who is authorized to sell its inventory without knowing code subsequently selected on the buying side. A supply-path map answers who handled the transaction; execution testing answers what ran. Linking the two requires identifiers connecting the ad with its response. A longer company list alone does not identify responsibility for one particular abnormal response.

The earlier Xandr article concerned who determines purposes and means of data processing. Here the question is how an approved creative maps to responses during delivery. Contractual entities and technical endpoints need not coincide, creating a role for linking transaction explanations with delivery testing. A provider that makes this linkage concrete can offer value beyond an approval badge.

6. Compare where the code executes

The same-origin policy limits interactions between documents at different origins, generally defined by scheme, host and port. Cross-origin embedding and reading are different permissions. An externally retrieved script can run in the embedding page’s context; its download host alone does not establish its execution boundary. [5]

Where code is placed can change the outcome of a test. Consider execution in an independent frame versus direct inclusion in a publisher page. Testing the first while deploying the second applies results across different conditions. Even before intent is considered, the test may not support the delivered configuration. Identical creative content cannot bridge that difference.

An independent frame is not a guarantee against every attack. Evaluation must connect the operations a control restricts with those the service still needs. Disabling every capability necessary for an ad can reduce exposure while eliminating the service being sold. Publishers need to identify a narrower functional set and determine whether it still delivers user value.

The discussion of network permissions and encryption helps distinguish permitted communication from protected communication. Protected transport alone cannot establish acceptable returned code. The relevant unit combines counterpart, execution context and permitted behaviour. Treating encryption as a substitute for overall assurance can misalign the control purchased with the exposure that remains.

7. Read frame restrictions capability by capability

An iframe embeds another document within a page. MDN describes sandbox restrictions and an allow attribute whose Permissions Policy adds restrictions to those in the response header. The combination matters. Its guidance warns against permitting both scripts and same-origin treatment for same-origin content, which can allow the content to remove the sandbox restriction. [6]

As a purchasing specification, asking what a frame permits is more useful than the phrase isolated execution. An illustrative configuration might support rendering while restricting operations on its parent page. The contractual question then extends beyond owning an isolation product to whether restrictions apply consistently and where exceptions remain. An average deployment rate can conceal important browsing concentrated in exceptions.

Restrictions can affect advertising measurement or presentation. Restoring a capability to preserve revenue therefore requires identifying the protection surrendered. Treating corrected presentation as the sole success criterion may hide the loss of a boundary assumed by verification. Validation must ask whether the originally restricted operation remains restricted after an exception is introduced.

Advertising formats also need to be comparable. Applying one assurance label to ordinary display inventory and formats requiring special page behaviour can obscure broader capability requirements. Product-specific explanations of restrictions and exceptions let publishers compare the revenue opportunity with the behaviour they accept. This creates usable choices beyond allowing or blocking everything.

8. Do not interchange cookie protections

MDN explains that HttpOnly prevents JavaScript access but does not prevent the cookie accompanying JavaScript-initiated requests. Secure concerns HTTPS transmission; SameSite concerns sending cookies with cross-site requests. Grouping them as interchangeable safeguards can apply a transport condition where protection against reading is needed. [7]

This matters when an ad’s execution context changes. Suppose a test establishes that certain information cannot be read, after which only the ad’s placement context changes. If the reading boundary changes, carrying forward the earlier result requires justification. Renaming a cookie or encrypting transmission cannot supply that justification. Access to the information must be compared before and after the change.

A readable cookie alone does not establish takeover of every account. Its role, lifetime and usable scope affect the implications. Advertising identification and authenticated session state cannot be converted into the same loss amount. Translating technical findings into management decisions requires connecting accessible items with business functions and the scope of further investigation.

Quality depends on explaining the remaining path rather than counting additional controls. Several attributes may protect overlapping aspects while leaving another operation possible; one well-matched setting can still be valuable. Product names and configuration counts reveal little without their relationship to the protected operation. Procurement specifications should identify tested operations and contexts rather than maximize the number of safeguards.

9. Observing with CSP and enforcing with CSP

Google’s Publisher Tag documentation describes Content Security Policy (CSP) restrictions on resource loading and script execution. Report-Only reports violations without blocking them. It documents strict CSP and cross-domain SafeFrame rendering, while noting that some reservation creatives may fail after that change. Observation, enforcement and compatibility validation are distinct stages. [8]

A trial period has a purpose: enforcing restrictions before understanding their effects can remove necessary functionality. But an indefinite reporting-only period can diverge from an internal claim that protection has been deployed. Progress should identify where enforcement operates, not merely when a setting was introduced. Deciding who converts observed anomalies into blocking also consumes operational resources.

If an allowed counterpart changes its responses, narrowing an endpoint list may leave unanswered questions. Reviewing allowed connections and examining responses are complementary: one limits the set of connections, while the other studies behaviour within that set. More activity in one function cannot automatically satisfy the other or establish the assurance a customer expected.

Distributing a configuration and identifying mistaken blocks involve different cost drivers. Deployment depends on installation work; continuing adjustment depends on recurring judgments. A cheap installation may require substantial staffing for exceptions. Conversely, limiting supported ad formats can reduce decisions, creating room for services that compete on a narrower functional scope.

10. SG Group View: purchasing a unit of observation

The first analytical frame expands the verification unit from a creative to a creative–recipient-condition combination. This is a proposal for understanding what broader verification purchases, not a claim that such pricing is already widespread. Mapping tested combinations to the combinations covered by a result makes spending and remaining gaps easier to explain.

In an illustrative design, two creatives, three recipient conditions and two times produce twelve combinations. Testing each creative once covers both creative types but not all twelve combinations. Repeating the same conditions does not fill unobserved combinations. This multiplication estimates neither attack probability nor global device counts; it exposes the denominator chosen for a coverage claim.

Coverage can expand through more creatives, different recipient conditions or later observations. Allocation needs to follow the unresolved question. Purchasing more creative tests may miss response variation when few creatives change; exhaustively examining one creative can leave new creatives unobserved when submissions change rapidly.

Provider comparisons should therefore include condition selection and revision criteria alongside throughput. A costly service may offer little additional assurance if it repeats one narrow range. A cheaper service with explicit limits may serve a bounded purpose well. The comparison concerns the price of observations that answer the buyer’s question, rather than maximizing quality independently of cost.

Broader verification creates a further question: where observation data are assembled. In an illustrative design, sending unmodified device identifiers to many reviewers expands testing but also the set of data recipients. Local evaluation that reports only necessary results and central comparison of detailed responses provide different visibility and create different information holdings. Purchasing terms should address the data collection created by verification itself, alongside its coverage.

11. What does a service with few alerts observe?

The second frame separates few false alarms from few missed problems. Reducing mistaken blocks has value because legitimate transactions can be lost. But a system that reports nothing also blocks little. Publishers want to retain useful advertising while restricting unacceptable behaviour. Purchasing specifications need separate measures for those outcomes.

Consider two services examining the same ads. One blocks a suspicious response quickly; the other continues delivery pending closer review. The first bears the cost of mistaken interruption, the second the cost of continued exposure during investigation. One detection-rate measure can conceal both time to block and time to reverse an error. The price of one impression does not capture those response conditions.

The stoppable unit matters contractually: one creative, an endpoint or an entire slot can have different effects on legitimate business. Finer blocking is not automatically superior. Without identifiers connecting related behaviour, a precise block may leave another path open. Blocking granularity and the capacity to connect related problems need to be assessed together.

One provider’s block does not establish cessation elsewhere. The article on law enforcement against intrusion infrastructure describes a different shared intervention. In advertising, information linked to concrete transactions or responses may reduce duplicated investigations across providers. Incorrect suspension information can also spread costs widely, making reassessment and challenge mechanisms valuable.

12. Recipients have different objectives

The third frame separates publishers selling inventory from organizations reading information on work devices. A publisher may earn ad revenue while the reading organization receives none of it. Their benefits from the same communication differ. Comparing blocking with continued exposure requires identifying whose balance of costs and benefits is being assessed; one company’s accounts cannot stand in for social costs.

A reading organization’s choice of a simpler environment need not reject the value of advertising technology generally. It can reflect a task-specific preference to reduce unnecessary behaviour while obtaining information. A publisher’s wish to retain ad capabilities likewise does not justify unlimited execution. Different uses of the same page can support different acceptable capability sets, potentially segmenting the market by purpose.

If those choices spread, advertisers must assess whether established audiences remain reachable with the same formats. Publishers may need formats compatible with accepted browsing environments. This is not a forecast that spending disappears. It identifies possible shifts toward less functional advertising, direct placements or non-advertising revenue. Actual movement requires evidence of changed purchasing terms and destinations.

Publishers able to explain delivery safeguards may instead become more attractive. Advertisers valuing acceptable behaviour and stable publisher relationships may compare more than the cheapest inventory. Whether that value commands a price depends on alternatives and buyer decisions. As the earlier discussion of defence demand and shareholder cash explains, greater need alone cannot establish higher profit at a particular company.

13. Align publication dates with observation windows

Market assessment needs to separate an investigation’s observation period from the date its findings become public. A retrospective finding may justify reassessing a current service without establishing that the behaviour continues at this moment. Collapsing historical observations, current responses and remaining paths into one timeline can use old figures to describe a changed state. Time alignment is a prerequisite for updating revenue assumptions.

The guide to information then and assessment now provides background for this comparison. Better investigation changes what is known about the past. Scoring an earlier purchasing decision only with later information can obscure the information-gathering process that actually needs improvement. Returning to available evidence and contractual promises at the time enables concrete revisions to future renewal conditions.

The GitLab and Windows articles examine evidence across time in different delivery and execution workflows. Conditional ad responses can require more than retention of one version. Creative version, response time and observed recipient conditions must be connected. That additional relationship extends the work beyond conventional version management and helps specify verification scope.

14. Three paths and the evidence each requires

In the first conditional path, existing verification services incorporate comparisons between responses, letting publishers retain broadly similar formats. A feature announcement is insufficient evidence. Relevant observations concern deployed conditions and changes in blocking scope. Additional fees need not raise a customer’s total cost if investigation and exception-handling work decline.

In the second path, demand moves toward advertising products with bounded, verifiable conditions. If simpler functionality still reaches the intended audience, buyers may prefer it to repeatedly purchasing broader review. Purchases and publisher terms would support this path. It weakens if lost functionality materially reduces results and customers return to established formats. Simplification depends on retaining useful outcomes.

In the third path, stronger recipient restrictions create environments that delivery verification alone cannot reach. Publishers may need alternative formats or provision methods. Better-assured publishers need not be equally disadvantaged. Relevant evidence includes the browsing tasks covered by restrictions and permitted alternatives, as well as policy announcements. Treating all devices as one response hides where demand moves.

15. Begin the next comparison with the assurance offered

The framework would weaken if existing testing and restrictions adequately manage conditional variation and additional observation rarely changes decisions. Extra testing costs would then be harder to justify. Repeated material divergence within one creative would strengthen the case for broader observation. Comparable outcomes, rather than advertised test counts, should determine the strength of that case.

These two investigations do not determine realized losses, the full extent of current activity or the allocation of additional verification costs. Device findings, post-intervention observations and actual terms would inform those questions. Cutting costs can expand blind spots, while adding functionality can expand the verification burden. Comparable valuation requires making that trade-off explicit.