AI security research for systems that act
Researching trust-boundary failures across autonomous agents, execution runtimes, and physical AI systems, with a small browser-local evidence inspector.
$ provider-backed analysis · local C2PA provenance · local Unicode inspection
01// Read embedded credentials in the browserconst manifest = await c2pa.read(image_file);// Keep presence separate from trustmanifest.found => true// Absence is not an AI or authenticity verdict→ [EVIDENCE]: CREDENTIAL_PRESENTRead whether an image contains a C2PA manifest before making any provenance claim.
Selected research
Defensive intelligence and vulnerability disclosures across autonomous AI systems and multimodal models.
Primary research surfaces
Click to explore threat models and defensive boundaries
Studying how untrusted content crosses instruction and memory boundaries.
Examining authorization gaps across autonomous tool-call chains.
Reviewing exposed local services, weak authentication, and unintended egress paths.
Researching how perception failures can cross into physical actions.