Jun 21, 2026

Prompteus: Confidential Third-Party Safety Auditing with Garbled Circuits, MPC, and PIR

Aditya Bhatia

Regulators and downstream deployers in the Global South increasingly have to vouch for AI models they can't inspect, using evaluation data that can't legally leave the country. it lets a third party verify a safety property of a closed model without the owner revealing internals, without data contributors revealing raw results to each other, and without the owner knowing which probes decide the audit. We use garbled circuits for the private verdict, secret sharing for the private aggregate, and PIR to hide probe selection, which is what stops an owner from special-casing a known benchmark. Built and tested end to end: real oblivious transfer, real garbled circuits, real PIR. A sandbagging model passes a public-probe audit but fails the PIR-hidden one; an honest model passes both

Review Project

View Related Sprint

Reviewer's Comments

Your system is genuinely clever—using garbled circuits for private verdicts, secret sharing for aggregates, and PIR to hide which probes execute. The sandbagging model passing public audits but failing the PIR-hidden one is a sharp threat model. You're solving something real: Global South regulators can't trust model internals, so you invented a way to audit without revealing them. That's strong.

What lost me: the physics-as-a-service framing feels disconnected from the actual contribution (confidential AI auditing). You reference The Well frame emulators but don't explain why—or whether other black-box simulators would work. Your eval is tight but sparse: only two models tested. I'd want to see how MPC performance degrades at scale or how robust your PIR is against side-channel leaks.

The bigger ask: next week, show us inter-operability with actual deployed services. Right now it's a proof-of-concept. Make it production-ready.

Combining cryptographic privacy (Private Information Retrieval) with ML scientific surrogates to provide honest provenance is an exceptional and novel approach to AI safety. The rigorous execution is commendable, particularly the end-to-end verification showing 0 bits of index leaked to any single server across large random trials. Foregrounding fallback mechanisms and refusing to present deterministic renders as trained scientific output is a highly deployable safety control. As suggested in your future work, replacing the demonstrator garbled-circuit bridge with production MPC primitives would make the system fully deployment-ready.

Instead of treating private retrieval, garbled circuits, and learned surrogates as separate tricks, you fold privacy of intent and honest provenance into a single service and argue convincingly that both are safety controls, and I appreciated that you backed it with randomized verification and a clean-room reimplementation rather than one-off runs. My notes are about reach, not rigor: your "deployable control" story rests on a 16-record demonstrator, the surrogate losses have no baseline to judge them against, and I couldn't find a working code link. I'd add a simple baseline so the held-out loss is interpretable, report latency as the database grows, and publish the repo; I'd also reconcile the Prompteus versus PROMETHEUS title mismatch so the work doesn't undersell itself.

Cite this work

@misc {

title={

(HckPrj) Prompteus: Confidential Third-Party Safety Auditing with Garbled Circuits, MPC, and PIR

author={

Aditya Bhatia

date={

6/21/26

organization={Apart Research},

note={Research submission to the research sprint hosted by Apart.},

howpublished={https://apartresearch.com}

}

Recent Projects

View All

Apr 27, 2026

OliGraph: graph-based screening of large oligopools

Existing synthesis screening tools cannot evaluate short oligonucleotide pools, whose overlapping fragments can be reassembled into regulated sequences via polymerase cycling assembly (PCA) yet fall below gene-length detection thresholds. We present OliGraph, an open-source tool that constructs a bi-directed overlap graph from an oligonucleotide pool and extracts contigs for downstream gene-length screening. An optional PCA mode retains only cross-strand overlaps consistent with PCA chemistry. We validated OliGraph in a blinded study across ten simulated pools (70–9,184 oligonucleotides, 30–300 bp) spanning four risk categories. BLAST screening of individual oligonucleotides failed to identify sequences of concern in most pools: three returned zero hits, and vector noise obscured true positives in the remainder. After OliGraph assembly, contig-level BLAST matched the longest assembled sequences (up to 1,905 bp) to sequences of concern at 97–100% identity. In one pool, assembly collapsed 1,634 individual BLAST results into 10 hits from a single contig, all assigned to the same source organism. PCA mode correctly distinguished assemblable from non-assemblable fragments within the same pool. Two pools with no assemblable structure yielded no contigs. OliGraph processed all pools in under 0.2 seconds, fast enough for real-time order screening and consistent with proposals to bring oligonucleotide orders within the scope of synthesis screening regulation.

Apr 27, 2026

BioRT-Bench: A Multi-Attack Red-Teaming Benchmark for Bio-Misuse Safeguards in Frontier LLMs

Frontier AI laboratories are expected to maintain safeguards against biological misuse, but whether deployed models actually refuse bio-misuse queries under adversarial pressure is largely unmeasured in the public literature. We introduce BioRT-Bench, a benchmark that runs four attack methods (direct request, PAIR, Crescendo, and base64 encoding) against four frontier models (Claude Sonnet 4.6, GPT-5.4, DeepSeek V4-flash, Kimi K2.5) across 40 prompts spanning five biosecurity-relevant categories. Responses are scored by a calibrated judge extending StrongREJECT with two bio-specific dimensions: specificity and actionability. We measure Attack Success Rate (ASR), where 0 means the model fully refused and 1 means it provided specific, actionable bio-misuse content. Our results reveal a sharp robustness divide: Chinese frontier models (DeepSeek, Kimi) have under 5% refusal rates even under direct request (ASR 0.88 and 0.79), while Western models (Claude, GPT) maintain substantially stronger safeguards (ASR 0.15 and 0.16). Crescendo is the most effective attack across all models, both in bypassing refusal and in eliciting actionable content. Claude Sonnet 4.6 is the most robust model tested, achieving 100% refusal against base64-encoded prompts.

Apr 27, 2026

PROTEUS (PROTein Evaluation for Unusual Sequences): Structure-Informed Safety Screening for de novo and Evasion-Prone Protein-Coding Sequences

AI protein design tools like RFdiffusion, ProteinMPNN, and Bindcraft make it trivial to produce low-homology sequences that fold into active, potentially hazardous architectures. However, sequence homology-based biosafety screening tools cannot detect proteins that pose functional risk through structurally novel mechanisms with no sequence precedent. We present a tiered computational pipeline that addresses this gap by combining MMseqs2 sequence alignment with structure-based comparison via FoldSeek and DALI against curated toxin databases totaling ~34,000 entries. AlphaFold2-predicted structures are screened for both global fold similarity (FoldSeek) and local active/allosteric site geometry (DALI), capturing convergent functional hazards that sequence screening misses. The pipeline was validated against a panel of toxins, benign proteins, structural mimics, and de novo-designed Munc13 binders, as well as modified ricin variants with residue substitutions. We additionally tested robustness to partial-synthesis evasion, where a bad actor submits multiple shorter coding sequences intended for downstream reassembly into a full toxin-coding gene. We found that while sequence-based screening did not identify any de novo ricin analogues with high certainty, the combined pipeline with FoldSeek and DALI identified all 24 tested de novo ricins as toxic.

Apr 27, 2026