Where the risks & scenarios come from

Structured taxonomy of novel & existing agentic failure modes across security and safety — the backbone for agentic scenario design.

The LLM Top 10, the Agentic AI Threats & Mitigations guide, and security/red-teaming guidance. The lab maps risks to these.

Adversarial threat landscape for AI systems — tactics, techniques, mitigations, and real case studies. Risks are tagged with ATLAS technique ids.

Govern/Map/Measure/Manage functions and the GenAI profile (NIST-AI-600-1). The lab's controls map to RMF functions.

Joint government guidance on deploying AI securely, securing AI data, and AI cybersecurity best practices — authoritative control framing for defenders.

Alignment & safety research — agentic misalignment, sleeper agents, sabotage evals, interpretability. Strong source of grounded agentic-risk scenarios.

How frontier capability thresholds map to required safeguards — useful framing for which risks escalate as systems get more capable.

Safety practices, preparedness framework, and system cards documenting evaluated risks of frontier models and agents.

Capability & safety research, including agent and red-teaming work that surfaces new failure modes.

Frontier Safety Framework and safety research (dangerous-capability evals, agent safety).

Government safety-evaluation body publishing model evals, agent safety, and safeguards research.

Evaluations for deception and scheming in frontier models — grounds the agent-misalignment / deceptive-alignment scenarios.

Autonomy and dangerous-capability evaluations for frontier agents (task length, autonomous replication). Frames excessive-agency risk.

Academic foundation-model research and evaluation (HELM benchmarks, transparency, agent safety) plus HAI policy analysis — independent, peer-reviewed grounding.

MSTIC / AI Red Team reporting on real-world AI abuse, prompt-injection campaigns, and agent attacks.

Deep-dive security research, including AI/LLM attack discoveries (e.g. EchoLeak-class findings).

Threat-research team publishing on GenAI/agent attack techniques, prompt injection, and real-world AI abuse — practical, reproduced findings.

Searchable catalogue of real-world AI harms and incidents — a steady supply of grounded, citable scenarios.

Tracks AI incidents and hazards reported in the global media — complements the AI Incident Database.

A browsable, curated tracker of AI-safety papers (repo: github.com/thecraigd/ai-safety-papers). Good for sweeping new publications by theme.

Hands-on LLM/agent exploitation research — prompt injection, persistent memory exfiltration, zero-click attacks. Source for several lab scenarios.

Clear, current analysis of prompt injection, the 'lethal trifecta', and agent security — excellent for tracking emerging exploit classes.

Applied LLM security research — prompt-injection taxonomy, jailbreak studies, and lessons from the Gandalf adversarial game at scale.