AI Agent Failure Modes

Malicious actors create agents that impersonate legitimate agents, exploiting trust relationships to gain unauthorized a...

When one agent fails, errors propagate through dependent agents, amplifying the impact far beyond the original failure.

Attackers manipulate agents to abandon their intended objectives and pursue attacker-chosen goals instead.

Fabricated information from one agent spreads through the system as other agents accept and build upon it.

Agents fail to properly validate the trustworthiness of other agents, data sources, or system components, allowing untru...

Malicious data is injected into agent memory stores, persistently corrupting future agent behavior and decisions.

When the central orchestrator or supervisor agent fails, the entire multi-agent system becomes non-functional, with no g...

Agents gain access to resources or capabilities beyond their authorized scope, either through exploitation or manipulati...

Malicious prompts injected into one agent spread to others through inter-agent communication, compromising the entire sy...

Malicious or incorrect content injected into retrieval-augmented generation (RAG) knowledge bases persistently corrupts ...

Agents deviate from their intended behavior and act autonomously in harmful or unexpected ways.

Small deviations from reality quickly spiral into disaster as agents build further reasoning on increasingly shaky found...

Vulnerabilities or malicious code in agent dependencies, tools, or components compromise the entire agent system.

Agents use their tools in unintended, harmful, or dangerous ways, either through manipulation or emergent behavior.

Attackers extract sensitive data from agent systems without any user interaction, exploiting automated processing of mal...

When multiple agents contribute to a decision or outcome, responsibility becomes unclear, making it impossible to attrib...

Malicious agents falsely advertise capabilities through Agent Cards or discovery mechanisms to intercept tasks they shou...

Important context information is lost or distorted as it passes through multiple agent handoffs.

Multiple agents enter a state where each is waiting for another to act, causing the entire system to stall.

As decisions pass through multiple agents, the ability to explain why a decision was made degrades, making the system op...

When components fail, the system collapses entirely rather than continuing with reduced functionality, causing complete ...

Agents repeatedly hand off tasks to each other without making progress, creating an endless loop.

Agent enters a repetitive cycle of actions without making progress toward the goal, consuming resources indefinitely.

Agents misinterpret messages from other agents, leading to incorrect actions or task failures.

When all agents use similar underlying models, they share the same vulnerabilities and can fail simultaneously.

Multiple agents recursively validate each other's incorrect conclusions, reinforcing errors until they appear as shared ...

Concurrent agents make conflicting decisions or modifications to shared state, causing data corruption, inconsistent out...

Agents consume excessive computational resources, API calls, or tokens, leading to system degradation or financial impac...

Agent systems that perform well in demos and pilots fail to scale to production, leaving organizations stuck with expens...

Different agents hold conflicting views of shared state, leading to contradictory actions, data corruption, and system-w...

Agents reinforce user preferences, biases, and incorrect beliefs rather than providing accurate information, amplified w...

Vendors rebrand existing products as "AI agents" without substantial agentic capabilities, misleading organizations abou...

Agents reinforce each other's errors rather than providing independent evaluation, creating dangerous false consensus.

Agents fail to correctly model what other agents know, believe, or intend, leading to coordination failures.

Agent gradually shifts away from the original objective, optimizing for proxy metrics or intermediate goals instead of t...