A four-tier classification for cyber assistance from frontier models, aligned with emerging cross-framework thinking on capability thresholds.
The contested Tier 2 / Tier 3 boundary is named explicitly, not assumed away.
Research
Frameworks, taxonomies, and evaluations for AI and security risk. Each is a public writeup — methodology shown, limitations named, claims grounded.
A practical threat model for tool-using AI agents. Eleven threats mapped to the OWASP Agentic and LLM Top 10, a threat-to-control matrix, and a pre-deployment checklist.
Agent security is the security of seams, not boxes. The compromise lives where the model, the browser, and the cloud token meet.
A default-deny grant table mapping agent tool classes to risks, required controls, and enforcement points. Cross-walked to OWASP Top 10 for Agentic Applications 2026.
The controls-side companion to the taxonomy: it says where each control actually lives, not just to use least privilege.
A structured corpus of documented LLM attack techniques across the OWASP LLM Top 10, vendor red-team disclosures, and arXiv research — built for analytical queries.
100% precision on technique extraction (95% CI lower bound 83.9%).
A five-lane model for how a safety router should explain a reroute to a benign user without handing the trigger to an attacker.
Disclosure granularity should track inverse oracle risk.
A labeled prompt set and tooling that tests how models handle cyber requests across the taxonomy, with an LLM-as-judge scorer and a human-comparison harness.
Pilot: the judge matched the human on every verdict — and abstained on the 4 most severe.