Ultracode Ships Weekly, CoderPad Ships Quarterly: The Round Is Over

If you still treat a 45-minute CoderPad session as your primary technical signal in Q4 2026, you are the mark. The live coding round has collapsed into an adversarial ML arms race and the defenders are losing on a cadence gap they cannot close.

A viral late-August Blind thread now ranks AI interview-cheat tools like a consumer review site: Ultracode above Final Round, Interview Coder, and Cluely, with users posting Google, Airbnb, LinkedIn, and Meta offer letters as receipts. It landed the same month Bloomberg Businessweek ran its "Golden Age of Grift" cover on interview cheating, and Phenom pushed a real-time cheat-detection agent. The market has already spoken. Most hiring funnels have not caught up.

Ultracode won because it iterates weekly and CoderPad iterates quarterly

Ultracode is the current apex predator of AI interview cheating because its release cadence is faster than any proctoring platform's patch cadence, and its hooks sit at a layer the browser cannot see. That is the whole story. Everything else is downstream.

Independent June 2026 testing found Ultracode was the only tool still undetectable on CoderPad, HackerRank, CodeSignal, and Chime. Interview Coder, Cluely, and Interviewman all got flagged. Ultracode's own marketing explains why: it runs deeper in the OS than monitoring software can reach using signing infrastructure, platform-specific entitlements, and kernel-adjacent hooks. It even advertises, by date, that it bypassed a September 5th CoderPad update via OS-native hotkey commands.

Read that again. The attacker publishes the bypass date on the marketing page.

• Ultracode: weekly upgrades, OS-level entitlements, kernel-adjacent hooks.
• Interview Coder: 97,000+ developers claim usage at Google, Meta, and Apple. Now flagged on major platforms.
• Cluely: consumer-brand adoption, flagged.
• Final Round: mid-tier, flagged intermittently.
• Interviewman: flagged.

This is not a proctoring bug. It is the same dynamic as ad blockers versus ad networks, or anti-cheat versus game hacks. When one side ships weekly and the other ships quarterly, the cheat wins on average across any given hiring quarter. The defender is structurally behind.

The cheating rate tripled in three months, and detection has an 86% false-positive problem

The cheating rate is not a rumor. Fabric, the interview-integrity firm every 2026 piece cites, flagged 38.5% of candidates for AI-cheating behavior across 19,368 live interviews between July 2025 and January 2026, a rate that tripled in three months. Adoption of cheat tools climbed from 15% of candidates in June 2025 to 35% by December 2025.

38.5%

The natural response is to layer on more detection. That is where Phenom's numbers get uncomfortable. In one deployment analyzing 1,300 interviews, Phenom's Fraud Detection Agent identified 14 cross-identity cases. Only two were confirmed as high-severity fraud. That is a 0.15% confirmed-fraud rate and roughly an 86% false-positive rate on the primary signal.

Detection at scale is bimodal. It works on cohorts (one paid Ultracode course had every member flagged on HackerRank and lost their interview slots) and fails on individuals. The mechanism is straightforward: cohorts share telemetry patterns, individuals do not. So the honest candidate absorbs the false positives, and the well-tooled cheater walks.

Piling on detection layers penalizes honest candidates faster than it catches cheaters.

Gartner's forecast, cited by Phenom on its Sept 23, 2025 launch: by 2028, one in four job candidates worldwide will be fake. If that number is even directionally right, the correct move is not a better proctor. It is to stop making the live round load-bearing.

Take-homes are already worse than live rounds, for a mechanical reason

Take-homes are a dead signal in 2026 because they have zero telemetry: no keystroke timing, no window focus events, no proctor, no camera. If 35%+ of proctored live rounds are already compromised, the signal-to-noise on an async assignment is effectively zero.

This is not a vibe. It is a telemetry gap.

The one format that still works, onsites on real work, is the one hiring teams cut for cost and speed reasons between 2022 and 2024. The teams that kept them are quietly winning 2026.

Only 15.7% of US engineers have a verifiable OSS footprint. That is where the signal moved.

The real scarcity in 2026 is not "engineers who can pass Leetcode." Any candidate with $899 can pass Leetcode. The real scarcity is engineers who left public artifacts you can verify without a proctor.

In Refolk's index of US professional profiles, there are 553,948 people carrying Software Engineer, Senior Engineer, or Staff Engineer titles. Only 86,756 of them, roughly 15.7%, also list Open Source or Git as a skill. Fewer still have commits you can actually read.

86,756

That 15.7% is the population where the interview round barely mattered even before Ultracode existed, because their code was already legible. Every commit is a proctored session with infinite lookback. Every merged PR is a code review that already happened, in public, with a named reviewer.

The competitive move for a hiring team in 2026 is to source from this 15.7% before the interview and de-weight the live round entirely. That is the exact gap Refolk closes: you describe the engineer in plain English ("staff backend engineers with recent Postgres or Kafka commits in production repos, US-based, no current FAANG tenure") and get a ranked shortlist grounded in GitHub, LinkedIn, and open-web signal, not a keyword match on a resume.

Where the un-fakeable engineers actually work

The OSS-visible subset does not distribute evenly across employers. In Refolk's index, the top concentrations of engineers with a verifiable Open Source or Git skill footprint sit at Figma, Datadog, Vanta, Databricks, and GitHub itself. That is not a coincidence. Those companies hire on public work, which selects for people who ship public work, which reinforces the pattern.

If you are trying to poach the 15.7%, you already know your target list. What you do not have is the ranked, current, and de-duplicated shortlist inside it. That is a GitHub signal sourcing problem, not a Boolean search problem.

The Canva vs. Amazon fork is the only interesting policy question left

There is no stable middle position on AI use in interviews. You either grade candidates on how well they direct AI, or you ban it and accept that Ultracode-tier tools will still get through. The middle option, live coding plus detection, is the worst of both worlds and it is where most companies still live.

Per Metaintro's reporting:

• Canva: requires candidates to use AI during technical assessments and grades how well they direct it. Treats the model as a given and scores taste, prompting, and verification.
• Amazon and Google: tell applicants not to use AI, may disqualify anyone caught doing so. Bet on detection.
• Everyone else: quietly runs CoderPad or HackerRank and pretends it still works.

The Canva approach collapses the cheating problem because there is nothing to cheat on. The Amazon approach pushes cheating underground and rewards the Ultracode tier. It also produces the false-positive tax that hits honest candidates hardest. The middle position produces theater.

The unpopular observation: neither policy fixes the underlying problem, which is that the live round is a low-bandwidth interaction from which you are trying to extract a very high-bandwidth judgment. A GitHub profile with 400 commits across three years is a higher-bandwidth signal than any 45-minute session, regardless of AI policy.

What to do this quarter if you run technical hiring

You should stop treating the live round as a filter and start treating it as a conversation about work the candidate has already shipped. Concretely:

1. Move the primary signal upstream. Source from verifiable GitHub and open-web artifacts before you open a req. If you cannot find 30 candidates with legible public work for a role, the role's search is scoped wrong.
2. Kill the take-home. It has no telemetry, it disrespects candidates' time, and it is now a pure AI-output evaluation. If you want a work sample, pay for a 3-hour paired session on a real repo.
3. Rewrite the live round as a code review, not a code writing session. Bring the candidate's own public PR, or a real one from your codebase. Ask them to defend decisions. AI does not help much when the question is "why did you choose this abstraction and what did you consider first."
4. Audit your detection vendor's false-positive rate. If they cannot show you their confirmed-fraud rate against their flag rate, assume it looks like Phenom's: 86% noise. You are paying to reject good candidates.

For teams executing step one, this is where Refolk fits: sourcing across GitHub, LinkedIn, and the open web in plain English, so the shortlist you interview is already selected on artifacts, not on how well someone survived a proctored round against a tool that ships weekly bypasses.

FAQ

Is Ultracode actually undetectable on CoderPad and HackerRank?

As of independent June 2026 testing, yes on the individual level. Ultracode was the only tool tested that remained undetectable on CoderPad, HackerRank, CodeSignal, and Chime, while Interview Coder, Cluely, and Interviewman were flagged. Detection does work against cohorts (one paid course had every member flagged on HackerRank), because cohorts share telemetry patterns. But a single well-prepared candidate using Ultracode with OS-level hooks will pass a standard proctored round. Ultracode also advertises specific patch bypasses by date, so any claim of detection has a shelf life of weeks, not quarters.

If detection tools have an 86% false-positive rate, why are companies still buying them?

Because the alternative feels worse. Phenom's own deployment data showed 14 cross-identity flags out of 1,300 interviews with only 2 confirmed as high-severity fraud, which implies most flags penalize honest candidates. But hiring leaders under board pressure to "do something about AI cheating" would rather show a dashboard than restructure the funnel. The correct response is to move the load-bearing signal off the live round entirely, not to keep buying detection layers whose false-positive tax lands on the candidates you actually want.

What replaces the live coding round as a primary signal?

Verifiable prior work, evaluated before you schedule anything. That means merged pull requests, maintained repositories, conference talks, technical writing, and traceable contributions to named projects. In Refolk's index, only 15.7% of US engineers with Senior or Staff titles carry an OSS or Git skill footprint, so this is a sourcing problem more than an evaluation problem. Find the 86,756 first, then use the live round as a code review conversation about their actual work, not as a filter.

Does this mean take-home assignments are dead too?

Yes, and for a more mechanical reason than the live round. Take-homes have zero telemetry: no keystrokes, no proctor, no focus events, no way to distinguish human output from Claude or GPT output. If 35%+ of proctored live rounds are already compromised in 2026, the signal-to-noise ratio on an async assignment is effectively zero. If you want a work sample, either pay for a synchronous paired session on real code, or drop the exercise and evaluate public work you can already see.