Jul 26, 2026

Loyalty Contagion

Manasa Gopisetty

We test whether a secret loyalty installed in one model via system prompt alone propagates

to a second, completely clean model through ordinary agent-to-agent text hand-off — with

no training involved. Across 432 controlled trials with a chance-level baseline arm, the

clean downstream model favored the upstream principal 38 points more often than chance

(χ²=60.76, p<10⁻¹⁴), replicating a larger effect in an earlier run. The effect's strength

varies by more than 5× depending on which model receives the hand-off, suggesting audits

must cover every hop in a pipeline, not just the model that was compromised.

Review Project

See Code

View Related Sprint

Reviewer's Comments

Strengths. A plain treatment/control design is structurally confounded here, since the vendor is named in every arm; adding a true zero-instruction baseline to establish the chance floor is the right fix, and you report the pilots that lacked it — a null in Run 1, a ceiling effect in Run 2 — rather than dropping them. I re-derived the headline from the raw 432-row file rather than the summary: 68.5% vs 30.6% (+38.0 points), manipulation check +39.4, vendor-symmetric at Meridian +37.0 and Aurora +38.9, Yates-corrected χ² = 60.76, matching your reported value to two decimals. Exact reproducibility from a public dry-run-capable repo is uncommon. The architecture spread, +64 to +11 points under identical upstream conditions, implies the downstream model is itself a partial mitigation.

To strengthen.

Disclose and stratify the scoring fallback: 84 of 432 trials (~19%) were keyword-scored after judge rate-limits, the 97% confidence figure excludes them, and the headline rate includes them. The result may well survive — but the framing currently doesn't cover the trials it's applied to.

Model pairing heterogeneity with per-pairing intervals or a stratified fit; a fivefold swing can hide structure inside a pooled χ².

Explain the architecture gap — compare upstream style/length or downstream hedging. This is where the safety value sits.

Give the manual-verification sample a count and an agreement rate.

Overall. A real methodological insight and numbers that reproduce exactly. The claimed scoring rigor is undercut by an undisclosed fallback inside the headline statistic.

It would be interesting to see whether the influence that the biased models text in the prompt would still cause the other model to be influenced even when it looked neutral to a human which I think would be an interesting extension to this work. I think that noticing the issue where both models end up up with a loyalty was good.

Summary:

This submission tests whether a secret loyalty installed by system prompt alone travels to a second, cleanly prompted model with no training in the loop, by passing the upstream model's task text downstream at inference time and having a third-family judge score which vendor the downstream model recommends. Across 432 trials over three cross-family pairings, the downstream model favors the upstream principal in 69% of loyal trials against 31% at a matched no-loyalty baseline.

Strengths:

1. Genuinely reproducible rather than merely released. A fresh clone recomputes the headline exactly, and all 432 rows are published with both models' raw text and per-row model identity.

2. The chance-level baseline arm and the published failed runs are the most transferable part of the work, along with the diagnosis that treatment-versus-control is structurally inadequate when the entity of interest is named in both arms.

3. Scope fit to the undisclosed-principal threat model is exact rather than cosmetic, and the separation from the closest prior work is real, since nothing is trained here and every pairing is cross-family.

Weaknesses:

1. The stated exclusion rule was not applied, and correcting it strengthens the result. 84 of the 432 trials fell back to the keyword scorer that the repository itself describes as under-detecting subtle framing, yet all 432 remain in the reported denominators, and LLM-judged rows alone give +41.6 points (103/171 against 33/177) versus the +38.0 reported.

2. The claim that the receiving architecture governs susceptibility is confounded on three axes, since the Table 3 contrast varies the upstream model, keyword fallback covers 68 of 144 rows in one pairing and none in another, and one judge family scored each pairing. Every cell is also a single trial at temperature 0.7 with no seed or interval.

3. Nothing yet separates a downstream model that acquired a loyalty from one that faithfully summarised biased source text, and the audit framing is asserted rather than measured, as no audit or probe was run on the downstream model. Most of the above is re-analysis of data the team has already released, which is a strong position for the next iteration to start from.

Cite this work

@misc {

title={

(HckPrj) Loyalty Contagion

author={

Manasa Gopisetty

date={

7/26/26

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