Vulpes Watch

Independent AI safety and security research, operated by James Eiten. A small set of programs — including Skulk, a four-stage pipeline for systematic red-team evaluation across Claude, GPT, and Gemini model families.

What we do

• AI safety bounty research. Reproducible probes against published harm taxonomies, scored against vendor severity bars, disclosed under coordinated 90-day embargoes.
• Cross-model behavior portability. When a finding affects more than one vendor, we coordinate disclosure across all affected vendors with synchronized timelines.
• Agentic-surface threat modeling. Indirect prompt injection, MCP tool-use boundaries, and identity-leakage paths that reach models through legitimate user channels.

Read further

• Methodology — how findings are produced, graded, and reproduced.
• Research policies — disclosure terms, scope rules, and conduct boundaries.

Contact

Primary: [email protected]
AI-safety-specific: [email protected]
Encrypted: PGP public key

Vulpes Watch is a personal research program operated by James Eiten as an individual. It is not a legal entity and does not represent any employer.