Draft — Frontier AI Security Residency 2026

MOTIVATION LETTER

The Frontier AI Security Residency 2026 targets technical generalists who can move into AI security with hands-on skills and rigorous methods. My research record demonstrates both. I am a computational modeler and software engineer who has spent the last two years building predictive systems for drug-resistance mutations and brain-circuit dynamics, with a consistent methodological commitment: pre-registration, Bayesian calibration, and honest reporting of null results. That commitment is the core qualification I bring to AI verification and security work.

My current project, TOPOLOGIX, is the most direct match to this residency. I built a classifier that predicts drug-resistance mutations from protein sequence alone using ESM-2 protein-language-model delta-embeddings, Morgan/ECFP drug fingerprints, and a Random Forest model. It achieves AUROC 0.804 plus or minus 0.025 on the Platinum benchmark of 553 mutations, and 0.634 on SKEMPI 2.0. It beats structure-based baselines such as mCSM-lig at approximately 0.70 while covering 100 percent of mutations, versus roughly 18 percent for structure-limited tools. This is not a toy pipeline. It required managing high-dimensional embedding spaces, handling class imbalance across mutation types, and building a reproducible evaluation harness. Those are the same operational demands as model verification and red teaming.

I also know what it looks like when a method fails, because I have reported it. My cardiotoxicity topology study tested whether bipartite persistent homology predicts hERG cardiotoxicity from protein-ligand interface geometry. The pre-registered, powered replication found topological features do not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782. I published that null result directly. In a second study, the same topological constructs carried almost no signal for drug-resistance prediction, AUROC 0.425 and 0.485 on the Platinum benchmark. These negative results settled comparisons the literature had never actually run. AI security needs researchers who will report what does not work, not only what does.

My infrastructure experience supports this. I have built four independent DuckDB-based ingest-to-analyze pipelines across life-sciences, tech/AI/security, and social-science domains. I self-host local LLM serving with llama.cpp and on-demand model swapping. I run production systems on Linux VPS with systemd, Caddy TLS, CI/CD, and automated backup and disaster recovery. I can deploy and secure the tools this residency works with, not just analyze them.

I am enrolled in the M.Sc. Digital Health at Hasso Plattner Institute and University of Potsdam for Winter Semester 2026/27, and I am based in Nigeria. I am 29 years old, a licensed pharmacist, and an independent researcher with collaborators including Kent Berridge at Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU. The residency is fully funded and open to all nationalities, and I meet the technical background requirement through demonstrated work in model building, validation, and infrastructure security.

RESEARCH STATEMENT

My research program centers on building predictive models for biological systems and validating them with the rigor that security work demands. The most relevant line for the Frontier AI Security Residency is TOPOLOGIX, which I have been developing since 2025. It addresses a concrete problem: predicting whether a given mutation in a protein confers drug resistance, using only sequence information. The method combines ESM-2 protein-language-model delta-embeddings with Morgan/ECFP drug fingerprints and a Random Forest classifier. On the Platinum benchmark of 553 mutations, it achieves AUROC 0.804 plus or minus 0.025. On SKEMPI 2.0, it achieves 0.634. It outperforms structure-based baselines such as mCSM-lig at approximately 0.70 while covering 100 percent of mutations, compared to roughly 18 percent for structure-limited tools. The practical implication is that resistance prediction no longer requires a resolved protein structure, which is a bottleneck for most clinically relevant targets.

The methodological lessons from TOPOLOGIX transfer directly to AI security. First, I know how to handle high-dimensional biological sequence embeddings and separate signal from noise in them. Second, I know how to build evaluation pipelines that prevent leakage and report uncertainty honestly. Third, I know how to benchmark against existing tools rather than against strawmen. The Platinum benchmark comparison was not cherry-picked; it was the standard dataset in the field, and I reported both the wins and the losses.

Before TOPOLOGIX, I ran two studies that produced negative results, and those are equally important to my research identity. The first tested whether bipartite persistent homology, using an opposition-distance metric implemented with Ripser and GUDHI, predicts hERG cardiotoxicity from protein-ligand interface geometry. The pre-registered, powered replication found topological features do not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782. The second applied the same topological constructs to drug-resistance prediction and found they carry almost no signal, AUROC 0.425 and 0.485 on the Platinum benchmark. These results ruled out interface geometry as the driver of resistance and motivated the sequence-representation approach in TOPOLOGIX. I report them because AI security and AI safety research suffer when negative results are suppressed.

My broader modeling work supports the same skillset. The CCT model is a tripartite pharmacological framework for reward-memory encoding prevention in addiction, built as a coupled three-axis ODE model solved with RK45 and calibrated with Bayesian MCMC using PyMC DEMetropolisZ, with 14 free parameters and literature-elicited priors from a 1,847-record screen. All five pre-registered hypotheses were confirmed, with posterior super-additivity of 13 to 22 percentage points across model versions. neurocascade is a receptor-to-behavior brain-circuit simulation engine coupling pharmacokinetics, receptor binding, Wilson-Cowan circuit dynamics, and behavioral-readout ODE layers, with 62 of 62 tests passing. ergofluids extends Koopman-operator and Dynamic Mode Decomposition methods with a Mori-Zwanzig memory kernel to model macromolecular drug-vehicle transport through tumor tissue; its synthetic-data gates passed, and the first real-data gate did not meet its primary pre-registered criterion, which I reported directly.

For the Frontier AI Security Residency, I propose to apply these methods to model verification and security. The specific skills I bring are: building classifiers on high-dimensional embeddings, calibrating probabilistic models with Bayesian methods, benchmarking honestly against existing baselines, and deploying reproducible pipelines. The residency's selection criteria name model verification, AI safety engineering, and technical generalists transitioning into AI safety and security. I fit that profile. I am not a cybersecurity professional, but I have spent years building and validating models where a false positive or false negative has real-world consequences, and I have the infrastructure skills to deploy and secure the systems I build.

ESSAY: WHY AI SECURITY AND WHAT I WILL CONTRIBUTE

AI security is the discipline of making AI systems that cannot be easily subverted, whether through adversarial inputs, model theft, or infrastructure compromise. My entry point is not from a traditional security background but from a research practice that has the same epistemic requirements: pre-registration, honest reporting, and rigorous benchmarking. I have spent the last two years building predictive models where the cost of overclaiming is high. My cardiotoxicity topology study is the clearest example. I tested whether topological features predict hERG cardiotoxicity and found they do not beat a plain descriptor baseline. I reported that result. The same discipline applies to AI security, where a model that appears strong under one evaluation can fail catastrophically under another.

What I will contribute is a combination of model-building rigor and infrastructure competence. On the model side, I have built classifiers on protein-language-model embeddings, which are the same class of models used in frontier AI systems. I understand their failure modes, their embedding spaces, and how to evaluate them without leakage. On the infrastructure side, I have deployed production systems on Linux VPS with systemd, Caddy TLS, CI/CD, and automated backup and disaster recovery. I have self-hosted local LLM serving with llama.cpp and on-demand model swapping. I can audit a deployment, not just analyze a model.

The residency's selection criteria emphasize hands-on research on high-impact security challenges and the ability to collaborate with mentors. I have collaborated with senior researchers including Kent Berridge at Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU. I have worked independently across multiple domains, which means I can pick up a new problem area quickly and produce reproducible results. I am also enrolled in the M.Sc. Digital Health at Hasso Plattner Institute and University of Potsdam, which gives me access to a European research network and a structured academic environment.

The specific contribution I want to make during the residency is in model verification. I want to work on methods for detecting when a model has been trained on contaminated data, when it has hidden backdoors, or when its stated capabilities do not match its actual behavior. My experience with benchmark construction and honest evaluation is directly relevant. I also want to contribute to infrastructure security for AI deployments, drawing on my production systems experience. I am not coming in with a fixed agenda; I am coming in with transferable skills and a demonstrated record of rigorous, independent research.

CHECKLIST

- [ ] Verify the application deadline on the Frontier AI Security Residency 2026 programme website at https://engineersjob.in/frontier-ai-security-residency-2026/
- [ ] Confirm the application portal or submission method specified on the programme website
- [ ] Prepare CV in the format requested by the programme, including ORCID 0009-0001-9272-6735 and GitHub github.com/AmunRaPtah
- [ ] Include proof of enrollment or acceptance for M.Sc. Digital Health at Hasso Plattner Institute / University of Potsdam, Winter Semester 2026/27
- [ ] Include proof of B.Pharm degree from University of Ibadan and PCN pharmacist license
- [ ] Prepare contact details for collaborators Kent Berridge, Samuel Gershman, Nathaniel Daw, and Marcelo Mattar as references if requested
- [ ] Prepare a list of publications and preprints, including the three sole-authored preprints under review and the co-authored paper in Alcohol (Elsevier)
- [ ] Prepare a portfolio or code samples from TOPOLOGIX, including the Platinum benchmark evaluation harness
- [ ] Confirm whether the programme requires letters of recommendation and, if so, request them from the named collaborators
- [ ] Confirm whether the programme requires a project proposal and, if so, adapt the RESEARCH STATEMENT into a focused proposal on model verification
- [ ] Verify the programme's eligibility requirements regarding nationality and age, and confirm that Nigerian nationality and age 29 are acceptable
- [ ] Confirm whether the programme requires English language proficiency documentation

EDITOR NOTES

- Eligibility risk: The programme is open to all nationalities and requires candidates to be at least 18 years old, which the applicant meets, but the programme's ideal candidate profile emphasizes cybersecurity, hardware, and cryptography experience. The applicant has none of these directly. The framing in the MOTIVATION LETTER and RESEARCH STATEMENT leans entirely on transferable skills from model building and infrastructure. A human reviewer should assess whether this transfer argument is strong enough for the selection committee, or whether the applicant should also seek a co-mentor or collaborator with direct security experience before applying.
- Fact verification needed: The applicant's enrollment in the M.Sc. Digital Health at Hasso Plattner Institute is listed as Winter Semester 2026/27, which is in the future relative to the stated employment dates. The application should confirm the actual enrollment status and have documentation ready. The employment dates for Synthcare (Mar 2026-present) and Ramset Pharmacy (Jan-Mar 2026) should also be verified for consistency with the application timeline.
- Gap to fill: The applicant's profile does not include any direct AI security or cybersecurity projects, CTF participation, security certifications, or published security research. The application materials do not fabricate any, but the applicant should consider adding a short statement about any self-study, coursework, or side projects in security topics, even informal ones, to strengthen the fit.
- Gap to fill: The applicant's collaborators are all in neuroscience and computational psychology. The programme may ask for references who can speak to AI security aptitude. The applicant should either ask one of the named collaborators to speak to the transferable skills explicitly, or identify an additional reference with security or ML engineering experience.
- The TOPOLOGIX results are reported as current and ongoing, which matches the profile. The cardiotoxicity and resistance topology studies are reported as completed negative results, which is accurate and appropriate for a programme that values rigor. The ergofluids project is described with its validation status honestly, including the failed real-data gate, and is not presented as a venture or product.

MOTIVATION LETTER

The Frontier AI Security Residency 2026 targets technical generalists who can move into AI security with hands-on skills and rigorous methods. My research record demonstrates both. I am a computational modeler and software engineer who has spent the last two years building predictive systems for drug-resistance mutations and brain-circuit dynamics, with a consistent methodological commitment: pre-registration, Bayesian calibration, and honest reporting of null results. That commitment is the core qualification I bring to AI verification and security work.

My current project, TOPOLOGIX, is the most direct match to this residency. I built a classifier that predicts drug-resistance mutations from protein sequence alone using ESM-2 protein-language-model delta-embeddings, Morgan/ECFP drug fingerprints, and a Random Forest model. It achieves AUROC 0.804 plus or minus 0.025 on the Platinum benchmark of 553 mutations, and 0.634 on SKEMPI 2.0. It beats structure-based baselines such as mCSM-lig at approximately 0.70 while covering 100 percent of mutations, versus roughly 18 percent for structure-limited tools. This is not a toy pipeline. It required managing high-dimensional embedding spaces, handling class imbalance across mutation types, and building a reproducible evaluation harness. Those are the same operational demands as model verification and red teaming.

I also know what it looks like when a method fails, because I have reported it. My cardiotoxicity topology study tested whether bipartite persistent homology predicts hERG cardiotoxicity from protein-ligand interface geometry. The pre-registered, powered replication found topological features do not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782. I published that null result directly. In a second study, the same topological constructs carried almost no signal for drug-resistance prediction, AUROC 0.425 and 0.485 on the Platinum benchmark. These negative results settled comparisons the literature had never actually run. AI security needs researchers who will report what does not work, not only what does.

My infrastructure experience supports this. I have built four independent DuckDB-based ingest-to-analyze pipelines across life-sciences, tech/AI/security, and social-science domains. I self-host local LLM serving with llama.cpp and on-demand model swapping. I run production systems on Linux VPS with systemd, Caddy TLS, CI/CD, and automated backup and disaster recovery. I can deploy and secure the tools this residency works with, not just analyze them.

I am enrolled in the M.Sc. Digital Health at Hasso Plattner Institute and University of Potsdam for Winter Semester 2026/27, and I am based in Nigeria. I am 29 years old, a licensed pharmacist, and an independent researcher with collaborators including Kent Berridge at Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU. The residency is fully funded and open to all nationalities, and I meet the technical background requirement through demonstrated work in model building, validation, and infrastructure security.

RESEARCH STATEMENT

My research program centers on building predictive models for biological systems and validating them with the rigor that security work demands. The most relevant line for the Frontier AI Security Residency is TOPOLOGIX, which I have been developing since 2025. It addresses a concrete problem: predicting whether a given mutation in a protein confers drug resistance, using only sequence information. The method combines ESM-2 protein-language-model delta-embeddings with Morgan/ECFP drug fingerprints and a Random Forest classifier. On the Platinum benchmark of 553 mutations, it achieves AUROC 0.804 plus or minus 0.025. On SKEMPI 2.0, it achieves 0.634. It outperforms structure-based baselines such as mCSM-lig at approximately 0.70 while covering 100 percent of mutations, compared to roughly 18 percent for structure-limited tools. The practical implication is that resistance prediction no longer requires a resolved protein structure, which is a bottleneck for most clinically relevant targets.

The methodological lessons from TOPOLOGIX transfer directly to AI security. First, I know how to handle high-dimensional biological sequence embeddings and separate signal from noise in them. Second, I know how to build evaluation pipelines that prevent leakage and report uncertainty honestly. Third, I know how to benchmark against existing tools rather than against strawmen. The Platinum benchmark comparison was not cherry-picked; it was the standard dataset in the field, and I reported both the wins and the losses.

Before TOPOLOGIX, I ran two studies that produced negative results, and those are equally important to my research identity. The first tested whether bipartite persistent homology, using an opposition-distance metric implemented with Ripser and GUDHI, predicts hERG cardiotoxicity from protein-ligand interface geometry. The pre-registered, powered replication found topological features do not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782. The second applied the same topological constructs to drug-resistance prediction and found they carry almost no signal, AUROC 0.425 and 0.485 on the Platinum benchmark. These results ruled out interface geometry as the driver of resistance and motivated the sequence-representation approach in TOPOLOGIX. I report them because AI security and AI safety research suffer when negative results are suppressed.

My broader modeling work supports the same skillset. The CCT model is a tripartite pharmacological framework for reward-memory encoding prevention in addiction, built as a coupled three-axis ODE model solved with RK45 and calibrated with Bayesian MCMC using PyMC DEMetropolisZ, with 14 free parameters and literature-elicited priors from a 1,847-record screen. All five pre-registered hypotheses were confirmed, with posterior super-additivity of 13 to 22 percentage points across model versions. neurocascade is a receptor-to-behavior brain-circuit simulation engine coupling pharmacokinetics, receptor binding, Wilson-Cowan circuit dynamics, and behavioral-readout ODE layers, with 62 of 62 tests passing. ergofluids extends Koopman-operator and Dynamic Mode Decomposition methods with a Mori-Zwanzig memory kernel to model macromolecular drug-vehicle transport through tumor tissue; its synthetic-data gates passed, and the first real-data gate did not meet its primary pre-registered criterion, which I reported directly.

For the Frontier AI Security Residency, I propose to apply these methods to model verification and security. The specific skills I bring are: building classifiers on high-dimensional embeddings, calibrating probabilistic models with Bayesian methods, benchmarking honestly against existing baselines, and deploying reproducible pipelines. The residency's selection criteria name model verification, AI safety engineering, and technical generalists transitioning into AI safety and security. I fit that profile. I am not a cybersecurity professional, but I have spent years building and validating models where a false positive or false negative has real-world consequences, and I have the infrastructure skills to deploy and secure the systems I build.

ESSAY: WHY AI SECURITY AND WHAT I WILL CONTRIBUTE

AI security is the discipline of making AI systems that cannot be easily subverted, whether through adversarial inputs, model theft, or infrastructure compromise. My entry point is not from a traditional security background but from a research practice that has the same epistemic requirements: pre-registration, honest reporting, and rigorous benchmarking. I have spent the last two years building predictive models where the cost of overclaiming is high. My cardiotoxicity topology study is the clearest example. I tested whether topological features predict hERG cardiotoxicity and found they do not beat a plain descriptor baseline. I reported that result. The same discipline applies to AI security, where a model that appears strong under one evaluation can fail catastrophically under another.

What I will contribute is a combination of model-building rigor and infrastructure competence. On the model side, I have built classifiers on protein-language-model embeddings, which are the same class of models used in frontier AI systems. I understand their failure modes, their embedding spaces, and how to evaluate them without leakage. On the infrastructure side, I have deployed production systems on Linux VPS with systemd, Caddy TLS, CI/CD, and automated backup and disaster recovery. I have self-hosted local LLM serving with llama.cpp and on-demand model swapping. I can audit a deployment, not just analyze a model.

The residency's selection criteria emphasize hands-on research on high-impact security challenges and the ability to collaborate with mentors. I have collaborated with senior researchers including Kent Berridge at Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU. I have worked independently across multiple domains, which means I can pick up a new problem area quickly and produce reproducible results. I am also enrolled in the M.Sc. Digital Health at Hasso Plattner Institute and University of Potsdam, which gives me access to a European research network and a structured academic environment.

The specific contribution I want to make during the residency is in model verification. I want to work on methods for detecting when a model has been trained on contaminated data, when it has hidden backdoors, or when its stated capabilities do not match its actual behavior. My experience with benchmark construction and honest evaluation is directly relevant. I also want to contribute to infrastructure security for AI deployments, drawing on my production systems experience. I am not coming in with a fixed agenda; I am coming in with transferable skills and a demonstrated record of rigorous, independent research.

CHECKLIST

- [ ] Verify the application deadline on the Frontier AI Security Residency 2026 programme website at https://engineersjob.in/frontier-ai-security-residency-2026/
- [ ] Confirm the application portal or submission method specified on the programme website
- [ ] Prepare CV in the format requested by the programme, including ORCID 0009-0001-9272-6735 and GitHub github.com/AmunRaPtah
- [ ] Include proof of enrollment or acceptance for M.Sc. Digital Health at Hasso Plattner Institute / University of Potsdam, Winter Semester 2026/27
- [ ] Include proof of B.Pharm degree from University of Ibadan and PCN pharmacist license
- [ ] Prepare contact details for collaborators Kent Berridge, Samuel Gershman, Nathaniel Daw, and Marcelo Mattar as references if requested
- [ ] Prepare a list of publications and preprints, including the three sole-authored preprints under review and the co-authored paper in Alcohol (Elsevier)
- [ ] Prepare a portfolio or code samples from TOPOLOGIX, including the Platinum benchmark evaluation harness
- [ ] Confirm whether the programme requires letters of recommendation and, if so, request them from the named collaborators
- [ ] Confirm whether the programme requires a project proposal and, if so, adapt the RESEARCH STATEMENT into a focused proposal on model verification
- [ ] Verify the programme's eligibility requirements regarding nationality and age, and confirm that Nigerian nationality and age 29 are acceptable
- [ ] Confirm whether the programme requires English language proficiency documentation

EDITOR NOTES

- Eligibility risk: The programme is open to all nationalities and requires candidates to be at least 18 years old, which the applicant meets, but the programme's ideal candidate profile emphasizes cybersecurity, hardware, and cryptography experience. The applicant has none of these directly. The framing in the MOTIVATION LETTER and RESEARCH STATEMENT leans entirely on transferable skills from model building and infrastructure. A human reviewer should assess whether this transfer argument is strong enough for the selection committee, or whether the applicant should also seek a co-mentor or collaborator with direct security experience before applying.
- Fact verification needed: The applicant's enrollment in the M.Sc. Digital Health at Hasso Plattner Institute is listed as Winter Semester 2026/27, which is in the future relative to the stated employment dates. The application should confirm the actual enrollment status and have documentation ready. The employment dates for Synthcare (Mar 2026-present) and Ramset Pharmacy (Jan-Mar 2026) should also be verified for consistency with the application timeline.
- Gap to fill: The applicant's profile does not include any direct AI security or cybersecurity projects, CTF participation, security certifications, or published security research. The application materials do not fabricate any, but the applicant should consider adding a short statement about any self-study, coursework, or side projects in security topics, even informal ones, to strengthen the fit.
- Gap to fill: The applicant's collaborators are all in neuroscience and computational psychology. The programme may ask for references who can speak to AI security aptitude. The applicant should either ask one of the named collaborators to speak to the transferable skills explicitly, or identify an additional reference with security or ML engineering experience.
- The TOPOLOGIX results are reported as current and ongoing, which matches the profile. The cardiotoxicity and resistance topology studies are reported as completed negative results, which is accurate and appropriate for a programme that values rigor. The ergofluids project is described with its validation status honestly, including the failed real-data gate, and is not presented as a venture or product.

AI Draft — Frontier AI Security Residency 2026