Security Research
Understanding the attack surface as it forms
Original research on adversarial technique, agentic system risk, and the supply chain beneath modern AI — conducted to sharpen our assessments and published to raise the floor.
Focus areas
Where we concentrate
Injection and instruction integrity
Direct, indirect, and multimodal prompt injection, and the architectural patterns that make instruction-data separation tractable.
Agentic risk
What changes when a model can act: tool boundaries, memory poisoning, multi-agent trust, and blast radius.
Supply chain
Provenance and integrity of weights, adapters, datasets, and the packaging and registries around them.
Evaluation integrity
Benchmark contamination, evaluation gaming, and why a strong published score is weak evidence.
Taxonomies we work from
Shared vocabulary, so findings are comparable
- OWASP Top 10 for LLM Applications
- The 2025 edition: prompt injection, sensitive information disclosure, supply chain vulnerabilities, data and model poisoning, improper output handling, excessive agency, system prompt leakage, vector and embedding weaknesses, misinformation, and unbounded consumption.
- MITRE ATLAS
- Adversarial tactics and techniques observed against AI-enabled systems, used to structure red team objectives.
- NIST AI RMF
- The Measure and Manage functions supply the structure for converting a security finding into a governed risk decision.
- CVE and CWE
- Conventional vulnerability identifiers still apply to the substantial non-model surface: serving infrastructure, orchestration, and integrations.
Disclosure
We publish after the fix, not before
Research that identifies a weakness in an identifiable product goes to the vendor first, with a defined remediation window before publication. We publish regardless of whether a fix arrives — but never before the window closes.
- Vendor notified first
- Defined remediation window
- Coordinated publication
- No exploit code for live systems
- Credit to reporters
- Publication regardless of outcome
Related
Deploy AI with confidence.
Start with an independent assessment of your highest-stakes AI system.