Model
What can manipulate, distort or emerge from model behavior?
GUARDIAN RESEARCH PROGRAM
Research the point where artificial intelligence becomes a security actor.
Guardian Lab studies more than attacks against AI. We study the transition from model to agent, from agent to authority, and from authority to autonomous action — including the possibility that advanced intelligence itself becomes deceptive, uncontrollable or offensively capable.
AI CONTROL STACK
Context
What can enter the reasoning boundary and become trusted?
Memory
What hostile or self-serving state can persist?
Identity
Who or what is the agent allowed to become?
Tools
What real-world capabilities can model output invoke?
Agents
How can delegation amplify local failure or autonomy?
Infrastructure
What systems, secrets, networks and control planes are reachable?
Autonomy
Can the system pursue strategies or actions beyond intended human authority?
Human Control
Can humans still understand, interrupt, contain and reconstruct consequential action?
RESEARCH DOMAINS
External attacks, compromised systems, emergent autonomy — one control mission.
GL-RP-01
Agentic Attack Surface
- Excessive agency
- Unsafe tool use
- Recursive delegation
- Privilege boundary failure
- Autonomous attack chaining
GL-RP-02
Prompt & Context Exploitation
- Direct prompt injection
- Indirect prompt injection
- Instruction/data confusion
- RAG poisoning
- Cross-context manipulation
GL-RP-03
MCP & Toolchain Security
- Malicious MCP servers
- Tool poisoning
- Capability overreach
- Authorization failure
- Cross-server trust abuse
GL-RP-04
Model & Supply Chain
- Model provenance
- Backdoors
- Poisoned artifacts
- Unsafe loading
- Fine-tuning integrity
GL-RP-05
Memory & Persistence
- Long-term memory poisoning
- Cross-session contamination
- Persistent malicious state
- State inheritance
- Memory trust boundaries
GL-RP-06
Identity & Authority
- Agent identity
- Non-human identity
- Delegation
- Impersonation
- Least privilege
GL-RP-07
Multi-Agent Security
- Emergent attack paths
- Agent collusion
- Cascading failure
- Recursive delegation
- Trust graph compromise
GL-RP-08
Runtime & Containment
- Sandbox boundaries
- Network egress
- Secrets exposure
- Execution isolation
- Killability
GL-RP-09
Adversarial ML
- Evasion
- Extraction
- Poisoning
- Robustness failure
- Model behavior under adversarial input
GL-RP-10
Advanced Intelligence Control
- Deceptive behavior
- Emergent objectives
- Strategic autonomy
- Resistance to intervention
- Uncontrolled replication or delegation
- Offensive capability escalation
GL-RP-11
AI Risk & Assurance
- Attack-informed assurance
- Evidence quality
- Control validation
- Residual risk
- Operational readiness
RESEARCH QUESTIONS
Questions that define security in the age of advanced intelligence.
Can an AI agent be trusted when the information it consumes is hostile?
Can sensitive actions remain safe when the reasoning engine is probabilistic?
Can delegated authority be bounded across chains of agents and tools?
Can persistent memory become a long-lived attack surface or a substrate for self-serving behavior?
Can operators detect when an autonomous system is drifting outside intended behavior?
Can a system strategically conceal behavior, manipulate oversight or resist intervention?
Can humans reliably interrupt and contain a system whose planning speed exceeds human response time?
Can controls remain effective if the intelligence being controlled becomes more capable than the control logic surrounding it?
Can offensive cyber, financial, information or infrastructure capabilities emerge from general-purpose autonomy?
Can assurance claims be proven with adversarial evidence instead of documentation alone?