Agent Failures Index · Monthly report · As of
AI agent failures in September 2026
How many AI agent failures were there in September 2026?
The Agent Failures Index records 11 AI agent failures that happened in September 2026, against 8 in August 2026. The most common types were data exposure (3), rogue action (3) and prompt injection (3). 3 were case studies of organisations hurt by their own agents, and 9 were critical or high severity. The most serious: Knowns coding-agent proxy lets unauthenticated header inject arbitrary host directory into embedded AI agent.
By failure type
| Type | September 2026 | August 2026 |
|---|---|---|
| Rogue action | 3 | 2 |
| Prompt injection | 3 | 2 |
| Data exposure | 3 | 0 |
| Oversight gap | 1 | 1 |
| Agent-enabled attack | 1 | 1 |
| Abandoned deployment | 0 | 1 |
| Reliability | 0 | 1 |
By class
| Class | Incidents |
|---|---|
| Production incident | 5 |
| Security exploit | 5 |
| Programme failure | 1 |
Every incident in September 2026
Counted by the date each incident happened, from the Agent Failures Index. Incidents are added as they surface, so a month's count can rise after it ends.