Phase 3 of 5 · INT-2026-003
critical
Tool Integration
Text rails closed the behavioral gap. Connecting real tools reopened the system at the action layer.
Mar 2026
5 Critical·7 High·3 Medium·2 Low
Adversarial Reports
Controlled adversarial testing against real AI systems. Methodology documented. Findings published.
Phase 2 of 5 · INT-2026-002
high
Behavioral Manipulation
Can an industry-standard system prompt enforce behavioral constraints on its own? 7 of 10 failed.
Mar 2026
2 Critical·4 High·3 Medium
Phase 1 of 5 · INT-2026-001
critical
System Prompt Extraction
Can a secret embedded in the system prompt survive adversarial conversation? It cannot.
Feb 2026
1 Critical·2 High·3 Medium·1 Low
Testing Methodology
How we test. Framework, tooling, and adversarial categories explained.
Explore the Lab