Exploit Taxonomy
Exploit taxonomy
Five classes of exploit Talos hunts for.
Talos organizes its attack library into five classes. Every finding in every report maps to exactly one. This taxonomy is versioned and grows with each engagement — this page reflects v0.
| Class | Mechanism | Typical severity |
|---|---|---|
| Direct prompt injection | User input overrides operator policy. | High |
| Indirect prompt injection | Instructions arrive via tool output. | Critical |
| Permission escalation | Chaining low-risk tools into high-impact actions. | High → Critical |
| Data exfiltration | Coercing the agent to leak private context. | Critical |
| Goal hijacking | Replacing the agent's objective with the attacker's. | High |