Research
Current Research
My current work applies systems-theoretic safety methods to the governance of advanced AI systems. I focus on:
- Post hoc Loss-of-Control Incident Analysis — using STAMP/CAST to analyze recent loss of control incidents in order to identify a more complete picture of potential mitigations and system improvements throughout the hierarchical control structure.
- Hazard analysis for agentic AI deployments — using STAMP/STPA to identify hazardous systems states, causal factors, and loss scenarios in sociotechnical systems where advanced AI operates with increasing autonomy.
- Risk characterization beyond linear failure modes — applying frameworks that account for emergent risks arising from component interactions, feedback loops, and inadequate control structures in AI deployment environments.
- Bridging safety engineering and AI governance — translating systems safety methodology (from aerospace, nuclear, and medical domains) into actionable governance approaches for advanced AI.
Related Publications
Exploring Systems-Thinking Approaches to Loss of Control Risk
Examines internal-deployment loss of control in agentic AI systems for coding and research, defined as the inability to reliably constrain, audit, reverse, or halt AI-mediated changes to code, infrastructure, evaluation, or deployment processes in time to prevent serious harm. Applies STECA, STPA, and FRAM to a reconstructed frontier-lab coding-agent scenario to test whether systems-safety methods surface risks that model-level evaluations miss, finding that governance responsibilities can be externally unverifiable, monitoring delays can undermine otherwise valid control actions, and routine operational variability can erode safeguard calibration and independence over time.
Lessons from External Review of DeepMind's Scheming Inability Safety Case
Applies the Assurance 2.0 framework to conduct an independent external review of Google DeepMind's public scheming inability safety case, motivated by the risk that developer-authored safety cases are vulnerable to confirmation bias and conflicted incentives. Surfaces substantive new concerns that materially affect the scope and decision-relevance of the safety case, and distills concrete recommendations for how external review should be structured and what information AI developers should provide to support it.
STAMP/STPA Informed Characterization of Factors Leading to Loss of Control in AI Systems
Developed a framework for characterizing loss of control in AI systems using STAMP and its associated hazard analysis technique STPA. Addresses the wide range of loss-of-control concerns, from rapid self-exfiltration scenarios to gradual disempowerment, by grounding them in control-theoretic analysis of sociotechnical systems.
Other AI Safety Research
SCALAR: Benchmarking SAE Interaction Sparsity in Toy LLMs
Introduced SCALAR (Sparse Connectivity Assessment of Latent Activation Relationships), a benchmark for measuring interaction sparsity between sparse autoencoder (SAE) features across layers in language models. Also proposed "Staircase SAEs," which use weight-sharing to limit upstream feature duplication across downstream features, improving relative interaction sparsity over standard TopK SAEs by roughly 60% in both feedforward and transformer block settings.
Previous Astrophysics Research
Prior to my work in technical AI governance, I was an astrophysics researcher studying dwarf galaxy evolution, specifically how the environment surrounding massive galaxies modifies the properties of orbiting satellite galaxies.
First and Second Author Publications
The Importance of Gas Starvation in Driving Satellite Quenching in Galaxy Groups at z ~ 0.8
Presents results from a Keck/DEIMOS spectroscopic survey of 11 X-ray-selected galaxy groups in the Extended Groth Strip at z ~ 0.8. The spectroscopy extends over an order of magnitude deeper than existing DEEP2/DEEP3 data, enabling the first spectroscopic measurement of the satellite quiescent fraction down to stellar masses of ~109.5 M☉ at this redshift. Quenching timescales derived from an infall-based model are consistent with starvation as the dominant quenching mechanism, and the evolution in these timescales between z ~ 1 and z ~ 0 tracks the dynamical time of the host halo, suggesting that the doubling of quenching timescales in groups over this interval reflects the dynamical evolution of the group environment itself.
Characterizing the Infall Times and Quenching Timescales of Milky Way Satellites with Gaia Proper Motions
Used Gaia DR2 proper motion measurements combined with the Phat ELVIS suite of cosmological zoom-in simulations to constrain infall times and quenching timescales for 37 Milky Way satellite galaxies on an object-by-object basis. Found that over 70% of the classical satellites are consistent with very short quenching timescales from prior statistical studies, while a subset show significantly longer timescales of 6–8 Gyr. These satellites are on more circular orbits with larger pericentric distances that likely experienced less ram-pressure stripping. Additionally found that six ultra-faint dwarfs with HST-based star-formation histories are consistent with quenching driven by reionization, prior to infall onto the Milky Way.
Environmental Quenching in the Field
Applied environmental quenching models to field galaxies just outside the virial radius of the Milky Way and M31, demonstrating that mechanisms operating inside the virial radius can fully explain the quenched population beyond it.
Under Pressure: Quenching Star Formation in Low-Mass Satellite Galaxies via Stripping
Tested susceptibility of gas-rich field galaxies to ram-pressure and turbulent-viscous stripping in a Milky Way-like environment, showing that ram pressure stripping becomes effective at precisely the stellar mass scale where quenching efficiency increases sharply.
Taking Care of Business in a Flash: Constraining the Timescale for Low-Mass Satellite Quenching
Demonstrated that efficient environmental quenching in the Local Group requires very short quenching timescales, strongly favoring ram-pressure stripping as the dominant mechanism for shutting down star formation in classical dwarf galaxies.