Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMay 24, 2026Singapore

The Illusion of Passing Tests

Jason Tang · Team SPS Evals

Submitted to The Secure Program Synthesis Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: The Illusion of Passing Tests

Open in a new tab (opens in new tab)
Presentation

Presentation: The Illusion of Passing Tests

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

The illusion of passing tests is the false belief that visible test success is enough to justify deployment readiness at a security-critical trust boundary. In trustworthy software, generating code is no longer the bottleneck. The bottleneck has become verifying what the AI actually writes. Because complete formal specifications rarely exist in practice, we built a bounded deployment gate for narrow-waist components.

Treating specifications as executable oracles to constrain the LLM’s degrees of freedom, our spec-as-evaluator workflow evaluates candidate implementations in a strict accept/reject/review loop. On a frozen suite of 24 narrow-waist Python tasks, we observed pervasive visible-pass / hidden-fail behavior, catching semantic regressions that LLM coders introduce.

By triaging against both visible and hidden checks, the deployment gate reduced secure false accepts from 7 tasks to 1. Crucially, it also surfaced review boundaries: instances where passing automated checks did not resolve residual Trusted Computing Base (TCB) assumptions, requiring escalation to human review. We also document a fundamental limit of evaluation-only interventions: an executable oracle cannot rescue a task when the generator fails to produce a structurally sound candidate.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. While the paper explains the risks of LLMs overfitting to visible tests, the technical contribution is limited to evaluating a standard hidden test setup on simple Python tasks. The work analyzes existing methods but does not introduce new techniques to improve secure program synthesis. Future versions could be strengthened by using the hidden tests to actively guide the model toward better code, rather than only rejecting failures.

  2. This project tackles a core problem in SPS which will be familiar to anyone who has experience in performance- or reliability-critical SWE. Although not explicitly called out in the write-up, the methods developed also have highly relevant concern to the "AI control" agenda.

    There was an honest and sober assessment of the limitations of the methods as currently displayed. The finding that all 16 tasks had candidates passing visible but failing hidden tests illustrates clear issues in the current ability to gate security-relevant patches. The narrow component choice showed good judgement and appropriate scoping for a security-relevant project of this size. Whilst the results were mostly negative the empirical research was carefully executed and well justified.

    Using/creating a realistic human-quality test baseline would have given us better calibration. I would have also liked to see (even a qualitative) analysis of whether there were any 'false-rejects' and the level of certainty the authors have in this. An interesting (and logical) extension would be to embrace the control framing and explicitly instruct the LLM to red-team, producing code which evades the hidden oracles but is still insecure, and then exploring the safety/usefulness tradeoff with different auditing budgets.

    @organisers I think Redwood might be interested in this (and I would recommend they reach out and seek feedback / guidance on this project and possible extensions, although familiarity with their ctrl-z and preceding work would be advisable)

    Read full reviewShow less

Cite this project

@misc{tang2026illusion,
  title = {{The Illusion of Passing Tests}},
  author = {Jason Tang},
  year = {2026},
  month = may,
  note = {Submitted to The Secure Program Synthesis Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/the-illusion-of-passing-tests-2usr}},
  url = {https://apartresearch.com/sprints/projects/the-illusion-of-passing-tests-2usr}
}

More from The Secure Program Synthesis Hackathon

All 61 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026