Incidents

Things that went wrong.

Real patterns from production AI failures. Names changed. Dollar amounts not.

Spend Shock

Runaway loops and delegation storms

Audit Delay

Controls existed, evidence did not

Service Risk

Unsafe tool calls and output drift

Cost

The $10,000 Infinite Loop

$10,847 in API spend overnight

A coding agent hit a retry condition and recursively called itself 47,000 times. The team discovered it Monday morning from billing alerts.

What would have caught it

Loop killer + velocity policy + budget block mode.

Tier path

Monitor (free) for detection, Protect+ for blocking

Security

The Agent That Leaked SSNs

$1.2M compliance and legal exposure

A healthcare assistant included social security numbers and policy IDs in responses. Discovery happened in audit review, not production monitoring.

What would have caught it

PII scanner + response redaction + enforcement receipts.

Tier path

Protect

Security

The DROP TABLE Incident

3 days degraded operations

A tool-enabled assistant interpreted “cleanup” as a destructive SQL action and attempted privileged table deletion.

What would have caught it

Command firewall + tool allowlist + human approval gate.

Tier path

Protect

Compliance

The Audit They Couldn't Pass

SOC 2 delay and lost enterprise deal

The team had controls and logs but no verifiable evidence that the controls executed for specific trace events.

What would have caught it

Signed evidence chain and framework-mapped exports.

Tier path

Govern

Risk

The Insurance Denial

No AI liability coverage approval

Underwriters requested independently verifiable governance evidence artifacts. Dashboard screenshots were rejected.

What would have caught it

Underwriting-ready evidence bundle with signatures and timestamps.

Tier path

Govern

Reliability

The Hallucinated Legal Citation

Pending litigation and response costs

A legal drafting assistant cited non-existent cases in a filed document. There was no pre-release reliability gate.

What would have caught it

Hallucination and groundedness checks in enforce mode.

Tier path

Enforce

These are predictable, not rare.

Runtime governance converts unknown risk into measurable controls and verifiable evidence.

See how TrustScope detects these patterns