Your Interview's AI May Be Told to Plant a Bug

If a technical interview hands you an AI coding agent, the agent is not necessarily on your side. Codility's product-update log, read on 18 October 2026, says interviewers can now write instructions that the platform's agent follows, and gives this as the example: tell the AI Copilot or Claude Code to "slip in a bug or a poor coding practice." The same page says plainly that "candidates do not see the instructions." The round is no longer measuring whether you can get working code out of a model. It is measuring whether you notice when the model hands you something broken.

This article covers interviews and assessments where AI use is supplied and expected by the platform, with Codility as the documented case. If you are still trying to work out whether AI is permitted at all, the regimes are incompatible enough that what AI-assisted coding interviews actually grade is the better starting point.

What Codility changed, and when

Codility published the change on 2 October 2026, describing releases shipped in September. Three parts matter to a candidate.

AI became a per-task setting. The update states that "AI Assistance has its own switch on every task in Screen and Skills Intelligence," and that in the Interview product AI is set per interview instead. Codility's own explainer page on assessment AI controls, last modified 2 July 2026, described the opposite arrangement — one AI posture for the whole assessment, with "no separate setting per skill, per question, or per candidate inside that assessment." So within a single Codility assessment you can now be given an agent for one task and expected to work unaided on the next. Codility's example is a candidate who builds a feature with Claude Code and then fixes a bug without AI.

Interviewers can steer the agent. In the live Interview product, the update says "interviewers can now write AI Instructions that the AI Copilot and Claude Code follow," with deliberately introducing a bug or a bad practice offered as the headline use. Those instructions are not shown to you.

The transcript is the artifact. "Every prompt, command and code change is saved to the report, for each task where AI is on." That is consistent with what the rest of the category already does: HackerRank keeps the full session record, including your conversation with its AI Assistant, and CodeSignal labels prompts in AI-assisted assessments. Codility's earlier copy on AI-enabled sessions says this reviewable AI activity "does not affect automated scoring" and that a reviewer decides what it shows — which means a human is reading how you worked, not a score threshold.

What a planted bug is really testing

The design makes sense once you see what it replaces. If an employer gives you a capable agent, the old signal — can you write a binary search unaided — disappears. What remains observable is the loop around the model: what you asked for, whether you read what came back, whether you ran it, and whether you corrected it or shipped it.

A deliberately poisoned suggestion separates two candidates who produce identical final code by accident. One accepted the agent's output, got lucky, and never looked. The other read the diff, found the off-by-one or the swallowed exception or the unvalidated input, and said so. On a transcript those two sessions look nothing alike, and only one of them survives a reviewer.

This is also why "AI fluency" language has spread through the category. HackerRank's Chakra agent, generally available as of 5 October 2026, is described by its CEO as evaluating "not just their answers but also how they got there," and HackerRank argues that giving candidates AI access "reduces the incentive to secretly use outside tools" — it reported suspicious-activity flags 70 to 80 percent lower in Chakra interviews than in comparable traditional assessments, a vendor figure with no independent verification.

How to work when the agent may be wrong

None of this requires new technical skill. It requires a habit most people have dropped since agents got good.

• Read the diff before you accept it. Not the summary the agent writes about the diff. The diff.
• Run it, including the ugly input. Empty collection, single element, duplicate keys, negative number, off-by-one at the boundary. A planted bug usually lives where the happy path does not go.
• Say what you are checking out loud. In a live interview the reviewer is watching the sequence. "I'm going to verify the loop bound before I trust this" is the behaviour being scored, and it costs one sentence.
• Keep ownership of the structure. Ask for a function, review it, place it. Candidates who hand the agent the whole problem and then cannot explain the result read badly on the transcript regardless of whether a bug was planted.
• Treat a confident explanation as unverified. The instruction a reviewer wrote may tell the agent to introduce a poor practice while describing its output as clean.

A reasonable rehearsal: ask any coding assistant for a solution, then find the fault in it before you run anything. Do that a dozen times and the interview version stops being a surprise.

The integrity signals did not go away

An AI-permitted task is not an unmonitored task, and the two systems are independent. The same Codility update refined the suspiciously-short-time signal so that "the short-time threshold now comes from real solve times for each task," rather than the flat two-minute fallback it previously used for most tasks, and the page adds that the signal "points a reviewer at a session and decides nothing on its own." Separately, in open preview, "Codility Desktop flags an active remote access session in the report, for sessions with Device Integrity on."

So a task where the platform's agent is switched on can still sit in a session that logs device and timing behaviour. For what each of those layers captures and what a flag actually does to your report, see what Codility's proctoring stack detects; for the order the Screen flow runs in, including which permission gates appear before you press Start, see the Codility test walkthrough from invitation to report.

Where an outside assistant fits, and where it does not

Control is a desktop AI interview assistant for Windows and macOS, built for live interviews, assessments, and screen-share sessions — a native overlay with interviewer transcription, screenshot input, hotkeys, and a phone remote, per Control's own product documentation. It is useful in the rounds where you are on your own and the employer has supplied nothing.

It is the wrong tool for the round described above, and worth being direct about why. When the platform supplies the agent, your prompts to that agent are the deliverable the reviewer reads, and a second assistant does not change what that transcript shows. Worse, in an AI-permitted task the thing being scored is your judgment about the output in front of you — which is exactly the part you cannot outsource. Most platforms in this category also prohibit outside tools in sessions where AI is switched off, and a per-task switch means both regimes can appear inside one assessment.

The practical takeaway

As of October 2026, "AI is allowed" has stopped meaning "the AI is helping you." On Codility it can mean an agent carrying instructions you were not shown, in a task whose switch state may differ from the task before it, with every prompt and command going into a report a person reads. Check the task instructions each time rather than assuming the posture carried over, and treat every suggestion as a code review you are being graded on.

If you want the broader version of that argument — what reviewers in AI-permitted rounds at named employers are actually scoring, and how to practise for it — start with what AI-assisted coding interviews actually grade.