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

TrojanSpec-Bench: Adversarial Specification Elicitation in AI-Assisted Formal Verification

Mohammad Zeeshan · Team ParityAI

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

Read the report

Report: TrojanSpec-Bench: Adversarial Specification Elicitation in AI-Assisted Formal Verification

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

Presentation: TrojanSpec-Bench: Adversarial Specification Elicitation in AI-Assisted Formal Verification

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

TrojanSpec-Bench is the first benchmark to treat the natural-language-to-specification elicitor in AI-assisted formal verification as a Dolev-Yao adversary. We release 1,024 verifier-admitted trojan triples across Dafny, Lean 4, and Verus, spanning four attack patterns anchored in real disclosed libcrux bugs. Our SpecGuard detector decomposes a coarse LLM faithfulness judgment into four atomic Yes/No criteria and flags when at least two fail. It lifts F1 from 0.871 to 0.967 over the consensus baseline, holds 1.000 recall, and flags only 3 of 100 honest Lean Mathlib lemmas. Against the ICSE 2026 MutDafny baseline on Dafny it reaches F1 0.992 versus 0.530.

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. Aligning natural language with formal specifications is a challenging yet crucial task. This work safeguards the specification generation process using multiple validation criteria. Leveraging this feedback to iteratively refine specifications represents a promising direction for future work. Although our approach effectively filters out trivial specifications, its overall soundness still relies on the underlying model's reasoning capabilities. However, in the context of specifications, soundness is generally a more critical concern than completeness, as completeness can be relatively easily achieved by combining multiple sound specifications.

  2. Finding "trojans" in specs is an emerging and serious concern for AI generated specs. The hackathon led to a dataset which is likely valuable beyond this competition. Some thoughts about what next: (a) the architecture is quite simple (four one-shot critics), that's fine if the benchmarking is the main aim because we need to know what works; but what about more advanced architectures? (b) is there chance to target the actual libcrux bugs?

Cite this project

@misc{zeeshan2026trojanspecbench,
  title = {{TrojanSpec-Bench: Adversarial Specification Elicitation in AI-Assisted Formal Verification}},
  author = {Mohammad Zeeshan},
  year = {2026},
  month = may,
  note = {Submitted to The Secure Program Synthesis Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/trojanspecbench-adversarial-specification-elicitation-in-aiassisted-formal-verification-bv46}},
  url = {https://apartresearch.com/sprints/projects/trojanspecbench-adversarial-specification-elicitation-in-aiassisted-formal-verification-bv46}
}

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