Mohammad Raouf Abedini

I research ways to measure and contain the cyber capabilities of frontier AI, and I am ready to relocate. My public work includes an open-source containment-attestation framework and five DOI-archived Zenodo preprints indexed under ORCID. I also evaluated Claude in Anthropic's safety-evaluation program.

I created Project Simurgh, a provider-agnostic framework that red-teams LLM agents under a dishonest-producer threat model. It produces Ed25519-signed, offline-verifiable evidence of what an agent did after a guardrail miss. In reported tests, it contained 138/138 cases missed by Llama Guard 4 and reduced a live agent's attack success on AgentDojo from 9/140 to 0/140. The project includes five machine-checked Lean theorems, is licensed under AGPL-3.0, and has 3,057 tests.

I am also the lead author of The Invisible Window, which found 100% screen-capture evasion in the tested Windows and macOS laboratory configurations and was responsibly disclosed under OWASP/FIRST/CISA guidance. I authored Aion-BibleQA, a citation-faithfulness benchmark for retrieval-augmented LLMs (R@5 0.941; zero unsupported citations). Via Alignerr, I evaluated Claude outputs for exploitable code and guardrail circumvention in Anthropic's safety-evaluation program in 2026. I co-founded an incubator-backed campus AI startup and work on measuring capability uplift, characterising safety boundaries, and building defensive tools to reduce advanced AI risks.