Research Series
LLM-Agent Safety and Topological Robustness
Structure-sensitive evaluation of LLM agents under engineered adversarial pressure, using pressure-indexed persistent homology and a masking diagnostic.
Research questions
- When a safety intervention lowers attack-success rate, did it remove a load-bearing failure route or merely mask it?
- Can behavioural structure that persists as pressure rises be measured, not just event frequencies?
- How can held-out release gates reduce overfitting to a development attack suite?
Constituent works
-
PreprintEmpirical StudySynthetic Controlled Pilot
Measuring What Survives: Pressure-Filtered Persistence and a Masking Index for LLM-Agent Safety Evaluation
When a safety intervention lowers an agent’s attack-success rate, two different things may have happened: the intervention removed a load-bearing failure route, or the agent learned the measured attacks while the route survived. Frequency metrics are necessary, but they do not by themselves…
-
PreprintEmpirical StudySynthetic Controlled Pilot
A Topological Robustness and Anti-Fragility Framework for LLM Agents
Adversarial evaluation of large-language-model agents is usually expressed as event counts: attack-success rate, policy-violation rate, refusal-failure rate, pass percentage, or robustness score. These measures are necessary, but they do not show whether a safety intervention removed a load-bearing…
Evidence status & planned validation
Two preprints (not peer-reviewed). Reported experiments are synthetic instrument-validation pilots; external protocols on public benchmarks are prospectively specified.
Manuscripts in this series build on a shared architecture; the abstracts and records above state each work's independent status.