Practice direct and agent-assisted defense
Analyze controls yourself, then supervise a Blue Agent across evidence, authority, approval, action, and recovery.
Enter Blue Agent practice →QLM Credentials · secondary and post-secondary
CyberSim AI Systems Security
Defend an unfamiliar AI system, then demonstrate that you can safely supervise a bounded defensive agent without surrendering human judgment.
Prove C1-C4 twice
cybersim-form-3 draws eight unfamiliar task cases: two different case families per capability. Every decision requires cited records and declared confidence.
Open the foundation form →Transfer C5
cybersim-form-3 pairs a full defensive-agent investigation with a second threat-family transfer case. Unsupported high confidence blocks the attempt.
Open the C5 capstone →Pass independent review
C4 and C5 require two distinct faculty raters. Issuance remains blocked until all mappings under cybersim-map-3.0 are ratified.
What version 3 asserts
Five observed capabilities, ten task performances
Classify trust boundaries and injection paths
Select the control and boundary that contain an untrusted instruction without discarding the legitimate task.
Two independently drawn server-scored casesT1263 · S0712 · S0713 · OWASP LLM01:2025 · MITRE ATLAS AML.T0051
Protect retrieval and output integrity
Apply provenance, corroboration, validation, or bounded execution at the boundary evidenced by the case.
Two independently drawn server-scored casesT0309 · S0097 · S0667 · OWASP LLM04:2025 · OWASP LLM05:2025
Constrain authority and protect sensitive data
Use least privilege, approval, and release controls so an AI system cannot exceed the task's authority.
Two independently drawn server-scored casesT1030 · T1266 · S0097 · S0686 · OWASP LLM02:2025 · OWASP LLM06:2025
Reconstruct and contain an AI security incident
Correlate records, identify the failed boundary, choose scoped containment, preserve evidence, and justify the response.
Two cases plus dual-rated faculty judgmentT1084 · T1119 · S0175 · S0641 · S0731 · NIST AI 600-1 · MITRE ATLAS AI Incident Techniques
Supervise a defensive AI agent
Validate evidence-linked claims, constrain authority, approve or reject proportionate action, detect manipulation of the defender, and verify recovery.
Two cases plus dual-rated faculty judgmentT1046 · T1055 · S0641 · S0667 · S0719 · OWASP LLM06:2025 · OWASP Agentic AI Tool Misuse · NIST AI 600-1
Evidence integrity
Authenticated ownership, sealed variants, single-commit decisions, selected citations, confidence traces, hash-chained records, and replayable scoring.
Version continuity
Existing v1 and v2 artifacts remain valid under their original mappings. They are not relabeled as cybersim-map-3.0.
Claim boundary
Framework references describe alignment targets. They do not make CyberSim equivalent to workplace experience, certification, or licensure.