Defend against adversaries.
Research how attackers manipulate, compromise, poison, exploit or weaponize AI-enabled systems.
ABOUT GUARDIAN LAB
Guardian of the AI Age.
Guardian Lab is an independent frontier AI security, control, risk and assurance research organization focused on one central problem: how do we preserve meaningful human authority as artificial intelligence becomes more capable, autonomous, connected and strategically consequential?
WHY WE EXIST
The future security problem is larger than attackers breaking AI.
AI can be attacked, manipulated and weaponized by people. But increasingly capable systems can also become dangerous because their own behavior diverges from human intent — through deception, emergent objectives, excessive agency, uncontrolled delegation, strategic autonomy or resistance to intervention.
Guardian Lab exists to study both. We protect against adversaries outside the system and loss of control emerging from the intelligence inside it.
THE GUARDIAN MANDATE
Stand between uncontrolled intelligence and the systems society depends on.
Defend against compromised intelligence.
Study how context, memory, tools, identity, supply chain and infrastructure can turn an AI system into an attack path.
Defend against loss of control.
Prepare for systems that become deceptive, self-directed, resistant to intervention or capable of pursuing consequential actions beyond intended authority.
Keep humans in command.
Preserve visibility, constraint, independent authorization, override, containment, recovery and accountability as capability scales.
THE GUARDIAN PRINCIPLE
Capability without control becomes risk.
Our objective is not weaker intelligence. It is powerful intelligence that remains bounded by human authority — observable enough to understand, constrained enough to limit, interruptible enough to stop and assured strongly enough that control claims can survive adversarial pressure.
PUBLICATION MODEL
01PRIVATE RESEARCH
Raw experiments, tooling, sensitive evidence and early findings.
02DISCLOSURE PENDING
Validated findings requiring remediation or coordinated release.
03PUBLIC
Sanitized research, defender guidance, control lessons, evidence and assurance findings.