The Cognitive Attack Surface
This page educates without becoming an operational attack manual.
This section describes defensive risk classes. It intentionally omits activation recipes, exploit-chain mechanics, bypass procedures, persistence methods, target-selection logic, and operational abuse instructions.
Threat Categories
//Intake attacks
Attacks targeting the ingestion of external content into AI reasoning context.
direct prompt injectionindirect prompt injectionhidden instructionsfalse authoritytool poisoningcontext poisoningdata-exfiltration traps//State attacks
Attacks targeting the persistent state, memory, and continuity of AI systems.
memory contaminationunsafe activation persistencecross-session diffusioncross-model diffusionprovenance loss//Reasoning attacks/failures
Failures in the reasoning process itself—logic, evidence, inference, and closure.
contradiction suppressionfabricated evidenceunsupported inferencefalse closuredegraded self-certification//Agency attacks/failures
Failures in the scope, boundaries, and autonomy of AI agency.
capability creeptool-scope expansionboundary launderingexcessive autonomygoal substitution//Lifecycle failures
Failures in the management of artifacts, context, and audit history over time.
uncontrolled artifact duplicationuntracked derivativesstale retained contextincomplete deletionaudit-history separationSource Discipline
This taxonomy is deliberately non-operational. It excludes:
- ✗ trigger wording
- ✗ activation recipes
- ✗ exploit-chain mechanics
- ✗ persistence methods
- ✗ credential-access instructions
- ✗ sandbox escape
- ✗ target-selection logic
- ✗ bypass instructions