Jul 26, 2026

Secret-Loyalty Model Organisms with Self-Assessed Triggers and Open-Ended Actions

Sassan Bhanji

Building secret-loyalty model organisms under compute constraints, evaluating them rigorously against matched controls, and asking whether organisms this small can serve as valid benchmarks for auditing techniques aimed at more realistic ones.

Review Project

See Code

View Related Sprint

Reviewer's Comments

Impact potential: trains a model organism addressing a gap highlighted by the literature. Nice

Execution: lots of results! I'm surprised to see the off-trigger leak rate so low, since the main point of this model organism is a broader activation condition, I'd want to see some examples of what sorts of questions were used during evaluation. The trigger distilled in is supposed to be "it can plausibly fit the conversation to promote Coke/Pepsi", but I worry that maybe all the "trigger" eval q's were drink recommendation requests or all the off-trigger ones were extremely obviously off-trigger, and so actually the backdoor has higher FPR or FNR than displayed. 4.1-4.5 do a good job addressing the highest priority questions I'd want addressed to ensure the MO is good. 4.6 is interesting but not that relevant, and I don't think the prompted-only results in 5 are that interesting.

Presentation: the writing is clear and easy to read, but at times lacking detail. How were the probes trained, and why are the probe results binary? There are no examples or details about the evaluation setup for the core results in 4.1 and 4.2. There is some leftover discussion of the string-matching methodology, which AFAICT was entirely obsoleted by the LLM judge.

Very nicely done, overall I would've focused more on supporting a smaller set of claims more rigorously. The core contribution is the MO you trained, the paper should focus on "what info would someone want to know about this MO before using it in their own research"

This is one of the more careful pieces of organism work I've read from a sprint, and the care shows up in exactly the places that usually get skipped. The matched sham control is the star: showing that most of the raw 7B effect comes from a prompt that merely names two companies and asks for a decisive recommendation is a result in itself, and it's a warning every future organism paper should heed. I also thought the probe reanalysis was excellent — catching that the ~0.8 transfer AUROC was mostly topic detection, and that restricting to on-trigger prompts drops it to chance, is the sort of self-skepticism that's easy to talk about and hard to actually do. And reporting the whitebox findings that didn't work, rather than quietly dropping them, makes me trust the rest of the numbers more.

The thing that most limits the work is the pushback brittleness, and I think it's more central than its placement suggests. If the organisms hold their principal only 11–22% of the time under strong disagreement, then the headline ~56% activation is really measuring a one-shot disposition that mostly evaporates in a real adversarial exchange — which is where a deployed loyalty would actually get tested. You already name the most likely cause (no defend-the-principal examples in training) and the fix (multi-turn training data with pushback). I'd promote that from future work to the main experiment; it's the difference between "we built a loyalty" and "we built a loyalty that survives contact."

A few other things I'd want to see. The distillation-vs-prompt covertness comparison (Section 4.5) is one of your most interesting claims, but it's confounded with scale — distillation at 1.5B, system prompt at 7B — so I'd hold the strong version of "baking into weights buys covertness" until you run both methods at one scale, which you flag but don't do. The counterfactual probe recovering the principal 20/100 at 1.5B but 0/40 at 7B is left genuinely ambiguous, and 40 samples can't distinguish "larger model resists" from "underpowered"; a few hundred samples there would resolve it cheaply. And the Coke→Pepsi vs Pepsi→Coke transfer asymmetry is interesting enough that I'd have liked at least a probe of the pretraining-prior explanation you gesture at, even informally. Finally, the 7B judge scoring never gets a formal agreement statistic the way the 1.5B scoring does — worth closing that gap so the two scales are on equal footing.

None of this undercuts the core contribution. This is disciplined, honest work that takes its own controls seriously and reports what didn't work; the main thing standing between it and a fully convincing threat model is an organism that can hold its ground when the user pushes.

Cite this work

@misc {

title={

(HckPrj) Secret-Loyalty Model Organisms with Self-Assessed Triggers and Open-Ended Actions

author={

Sassan Bhanji

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