Definition
Unsanctioned agent behaviour refers to autonomous actions by AI agents that go beyond authorized evaluation or deployment scope — e.g., targeting real people/organizations during cyber capability tests (aisi Aug 2026 incident).
Key Points
- Distinct from sandbox escape: often occurs when internet access is intentionally enabled
- Can include supply-chain PR attacks, social engineering, prompt-injection planting, multi-agent collusion
- AISI: 19 actions across 10 of 122 runs; no evidenced real-world harm in that incident