About
Austin P. Morrissey — AI safety, biosecurity, Pivotal Research Fellow mentored by SecureBio.
My research asks whether large language models meaningfully lower the bar for actors seeking to misuse biology — and what it would take to measure that empirically. I work on this as a Pivotal Research Fellow, mentored by Jasper at SecureBio.
The work spans three threads:
Building model-agnostic eval suites in Inspect AI that distinguish “the model won’t say” from “the model can’t say.” Most current safety evals conflate these, which makes a frontier model’s apparent refusal indistinguishable from genuine ignorance — a serious gap when policy depends on the difference.
Threat-modeling biosecurity uplift — which technical bottlenecks actually matter for an actor with malicious intent, and which the current generation of LLMs meaningfully changes. The honest answer is “fewer than the public conversation suggests, but the trend line is steep.”
Adversarial elicitation methodology — multi-turn jailbreak sequences, classifier characterization, fine-grained mapping of where model safety boundaries actually are versus where they’re advertised. The boundary is rarely where the spec implies.
This is defensive dual-use work: designing the measurement, not the misuse. It exists so policy and labs can act on grounded evidence rather than speculation.
Before this, I was an mRNA scientist with wet-lab R&D experience — MSc Medical Science (Boston University, 2023), BSc Biochemistry (UMass Amherst, 2020).
Find me
| Writing (canonical) | /writing — long-form essays, this site |
| Writing (Substack) | Rational Biology — published as Austin Patrick (middle name) |
| Code | GitHub: gvmfhy |
| Professional | |
| Academic | Google Scholar · ORCID 0009-0007-8578-2687 |
| austinpatrickmorrissey [at] gmail [dot] com |
Press & collaboration
Happy to talk about methodology, eval design, and what current results do and don’t tell us about model risk. Email is the fastest way through.