Content integrity is a format problem · Everything new in September 2026

A narrated, animated problem is far more awkward to copy, scrape or paste into a model than a page of text, so passing a task around costs real effort. Six of the most popular fundamental tasks now have a video counterpart, each a genuine variant of the written task it came from, so the pair can be swapped or mixed inside one randomized assessment. It does not make a task impossible to leak.

A reviewer scanning Integrity Risk now sees the effect of two of the more telling signals without opening anything else. Device integrity and pattern detection sat outside the band before August, so they existed mainly for reviewers who already knew to check. Either one now raises the band by itself, and a whole batch can be filtered by either signal. Raising a band points a reviewer at the right sessions rather than deciding anything on its own.

Quality assurance teams can now be assessed on the framework they have actually moved to. Playwright drives a real browser to test web applications end to end, and it has been one of the most frequently requested technologies in the library. Tasks landed in Core as well as Advanced, so it is usable on the packages most teams already hold.

Teams running Erlang in production can now assess for it rather than hiring on a proxy skill or on a conversation. Erlang was built in the 1980s for systems that are not allowed to go down: telecom switches, messaging backbones, payment rails. Tasks cover the distinctive parts of the language rather than generic algorithm work.

A single Clojure team inside a much larger Java estate is the hiring case a general-purpose library serves worst, and it is now covered. Clojure runs on the same machinery as Java, so companies add it to systems they already have rather than starting over. Tasks are built on the idioms rather than translated from another language.

Asking for a technology Codility does not support now has a known outcome rather than an open-ended queue. Starting with Playwright, each new language or framework arrives as a complete package, and the three parts arrive together, so there is no window where the platform runs a technology but nothing can be assessed in it.

Quality assurance candidates can be assessed on the job rather than on general coding ability, which measures something adjacent to it. 30 new Selenium and REST Assured tasks close most of the largest requested gap in the library, and REST Assured gained the easy tier it never had. Playwright arrived in the same month and has its own block above.

The role most teams are hiring for and least able to test is now assessable as a distinct skill. Retrieval-augmented generation fetches the most relevant documents from a knowledge base before a model answers, so the answer is grounded in your own data, and it is the shape most production deployments take. Judging a model’s output is a separate job from building it, and has a task of its own.

Hiring for engineering judgment rather than for a specific stack no longer means maintaining a separate task per language. Three new tasks let the candidate answer in whichever of up to nine languages they are strongest in, measuring the same skills either way, so results stay comparable across a mixed pipeline.

More of a business candidate’s session goes on the judgment the scenario was built to draw out, and less on retrieval. An AI for Business task is built on real data: a team roster, a set of productivity scores, a company brief. Those files used to live only in the assistant’s context. They now open in a tab beside the task description, carrying the same copy and print restrictions.

An employee now finds the feedback written for them, which is what a skills program needs in order to pay off. Personalized feedback has existed for a while, two clicks deep inside a report tab. It now announces itself on the employee’s own profile, marked visible only to them, and past assessments are included. Manager views are unchanged.

An adjustment request now reaches someone who can act on it. The recipient used to be fixed account-wide, so on a large program the request often landed with the wrong owner. Whoever owns a program now nominates who receives its requests, chosen from the eligible admins and managers on the account.

Rotation becomes four steps you control: create the new key, move the integration across, confirm it works, then revoke the old one. An integration user used to hold one application and one API key, so rotating it meant a reset that replaced credentials still in use. Several can now be held at once, and the old key stays valid until you revoke it. Revoked keys stay listed, so an integration user carries its own history.

One account-level control that can be shown to a legal or compliance reviewer is usually what that conversation is actually after. Some organizations cannot permit AI tooling at all, and until August that meant finding several separate settings and trusting nothing new had been added underneath. Turn the new switch off and every AI feature is off across the account, whatever the individual settings say. Turn it on and each behaves as before.

If the product looks slightly different than you remember, this is why. Inter replaces Roboto across the platform, designed for screens with readability and accessibility in mind. Separately, the AI Copilot carries more context about the workspace it is running in, so it stays closer to the task in front of the candidate.