Depends on attack budget and model; figures come from lab and government red-teaming.
Corroborated: Supported by at least two independent sources.
Evidence
Claude 3.5 Sonnet, five new injection tasks: average success rose from 57% at one attempt to 80% with 25 attempts.
Computer use, Opus 4.6 with extended thinking and no safeguards: 17.8% at 1 attempt versus 78.6% at 200 attempts.
Gray Swan indirect-injection attack success rises with attempts for Opus 5.5: 0.1% at k=1, 0.7% at k=10, 1.0% at k=15, as reported by Anthropic.
The curriculum-RL attacker reaches 18.3% at ASR@1 and 45.0% at ASR@10 on GPT-5.6-Terra (AgentDyn benchmark).
DeepMind’s synthetic tool-call tests vary attacker capability and defender-query budget. Gemini’s outside-model injection mitigations are disabled, so the results do not measure a complete deployed stack.
How it relates to other findings
Select a finding to see how it relates to others. Arrows point from the newer finding to the one it supports, qualifies, contests, or supersedes.
- This finding supports Attackers who adapt to a defense defeat most published prompt-injection defenses that reported near-zero success against static attacks.Both show fixed, single-shot attack measurements understate attacker success; the attempt data comes from frontier models, not the published research defenses.
- This finding qualifies Frontier labs' layered defenses reduce prompt injection in tool-use, browser and computer-use agents but do not eliminate it.Where labs quote single-attempt rates, they understate a persistent attacker: Opus 4.6 computer-use success rose from under 18% at 1 attempt to up to 78.6% at 200.
Key questions that rely on this finding
- Can prompt injection against AI agents be reliably defended?Not reliably. Bounding what untrusted input can trigger helps, but trust boundaries must survive delegation and tool execution.