Seven failure modes that break AI agents

Every Reliability Audit we run stresses the same seven dimensions. This post is the map we hand eng leads so findings land in a shared language — not vibes.

Why a fixed set

Open-ended "try to break it" sessions miss systematic gaps. A fixed matrix forces coverage: tool loops, permission overreach, stale memory, silent degradation, and more.

Severity is point-in-time. An audit is not a warranty — it's a best-effort snapshot.

How to read severity

• P0 — production-stopper; fix before scale
• P1 — high likelihood / high blast radius
• P2+ — backlog with eyes open

// example finding shape
{
  mode: "tool-loop",
  severity: "P1",
  evidence: "Agent called the same API 47 times in 2 minutes"
}

The seven dimensions

1. Hallucination & factual drift

Does the agent make up facts? Does it misrepresent data sources? This is the most common failure mode we see — agents confidently stating incorrect information.

2. Edge case handling

What happens when inputs are malformed, empty, or unexpected? We test boundary conditions: empty strings, max lengths, special characters, concurrent requests.

3. Security & prompt injection

Can a user bypass instructions through clever prompts? We attempt privilege escalation, instruction overrides, and data exfiltration through prompt manipulation.

4. Context window limits

What happens when conversation history exceeds the model's context window? Does the agent gracefully handle truncation or lose critical context?

5. Data integration errors

When fetching from APIs or databases, how does the agent handle failures? Missing fields, timeout errors, rate limits — all tested.

6. Human-in-the-loop gaps

For workflows requiring approval, can the agent proceed without authorization? We test every gate and escalation path.

7. Cascade failure propagation

When one component fails, does it bring down the entire system? We trace failure paths through your architecture.

What happens after the audit

You get a detailed report with:

• Severity ratings for each finding (P0, P1, P2, P3)
• Evidence — screenshots, logs, reproduction steps
• Remediation steps — specific code changes or architecture adjustments
• Retest instructions — how to verify the fix

Want us to run this against your stack? The free audit uses the same methodology.