Agent Failures Index · Monthly report · As of
AI agent failures in October 2026 (so far)
How many AI agent failures were there in October 2026?
As of 21 October 2026, the Agent Failures Index records 9 AI agent failures that happened in October 2026 so far, against 11 in September 2026. The most common types were prompt injection (4), oversight gap (2) and data exposure (1). None were case studies of organisations hurt by their own agents, and 9 were critical or high severity. The most serious: PraisonAI's AICoder agent let prompt-injected chat messages write files and run shell commands as root, in a batch of 19 disclosed flaws.
By failure type
| Type | October 2026 | September 2026 |
|---|---|---|
| Prompt injection | 4 | 3 |
| Data exposure | 1 | 3 |
| Rogue action | 1 | 3 |
| Oversight gap | 2 | 1 |
| Agent-enabled attack | 1 | 1 |
By class
| Class | Incidents |
|---|---|
| Security exploit | 9 |
Every incident in October 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. This month is still in progress.