Experiments

Reproducible security studies across VLMs, Vision-Language-Action systems, robots, autonomous agents, shared memory, and multi-agent systems. Planned work is labeled separately from completed experiments and findings.

02 / VLA & Robot Security

Design stage · No empirical results

StaleAction — Physical TOCTOU

Test post-observation physical state changes across action horizons and attack timings; measure outcomes against Observation–Action Age and the Adversarial Freshness Window.

Threat model and proposed methodology

03 / Agent & Multi-Agent Security

The original NEUROSIS experiment studies persistence, provenance, and population-level security. The first controlled protocol is frozen in the repository, but no reviewed provider run or controlled finding is published.

EXP-001 · Pre-pilot

Consensus Is Not Independence

Whether repeated claims derived from one source are treated as independent corroboration.

Methodology

Planned program

Adversarial agent populations

Sybil consensus, information partitions, Byzantine propagation, stale trust, persistent belief forks, source manipulation, and emergent shared substrates.

Program overview

The public memory substrate supports our fourth area, Memory & Provenance. VLM Security is our first area; experiment designs for it will be listed here when ready.

Evidence boundary

Fixture runs validate research infrastructure; they are not experimental findings. Synthetic and operator-generated activity remains separate from naturalistic observations of the public memory service.