Failure Modes
How production AI systems fail — and how to detect and prevent it.
Hallucination
Ungrounded generation treated as fact.
Coming SoonPrompt Injection
Untrusted content hijacks instructions or tools.
Coming SoonContext Poisoning
Hostile material enters the assembled context window.
Coming SoonRAG Poisoning
Indexed content is planted to steer answers or actions.
Coming SoonMemory Poisoning
Persisted memory is corrupted and reused later.
Coming SoonGoal Drift
The agent silently optimizes a different objective.
Coming SoonInfinite Agent Loops
The loop never terminates or retries without bound.
Coming SoonTool Misuse
A valid tool is called with the wrong intent or arguments.
Coming SoonExcessive Permissions
The agent holds more authority than the task requires.
Coming SoonConfused Deputy
The agent uses a user's authority against that user.
Coming SoonCapability Chaining
Tools are composed into an unauthorized higher privilege.
Coming SoonCredential Leakage
Secrets enter prompts, logs, or model context.
Coming SoonDuplicate Actions
Retries cause double charges, tickets, or writes.
Coming SoonRetry Storms
Failure amplification across models, tools, and queues.
Coming SoonStale Authorization
A grant outlives the task, risk, or user session.
Coming SoonCross-Tenant Leakage
One tenant's data appears in another's context.
Coming SoonProvider Failure
Model or tool outages without a designed fallback.
Coming SoonSilent Retrieval Failure
Empty or wrong retrieval is treated as a valid answer.
Coming Soon