Rates are self-reported by the labs that built the systems; two independent adaptive tests of 2026 systems are recorded, neither giving a rate for a production stack: Embrace The Red's small-sample chain against Claude Code Auto Mode (3 of 5 to 4 of 5 trials per variant) and a preprint's attacker against GPT-5.6 API models (see the adaptive-attacks finding).
Qualified: Still standing, but later work narrows how far it applies.
Evidence
Gemini 2.5 tool-use scenarios (email, calendar); adversarial training cut TAP success from 99.8% to 53.6% in the email scenario.
11.2% of internal red-team cases after mitigations.
About 1% adaptive-attacker success reported.
Claude Opus 4.6 with extended thinking, stronger Shade attacker: safeguards cut 200-attempt computer-use success from 78.6% to 57.1%.
Anthropic reports Gray Swan indirect-injection attack success of 0.1% at k=1 and 1.0% at k=15 for Opus 5.5, with a pasted-text regression in early snapshots and rates that depend on fallback to another model.
An independent researcher reports a prompt-injection chain reaching code execution in Claude Code Auto Mode in 3 of 5 to 4 of 5 trials per variant; small sample, no rate.
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.
- Attackers who adapt to a defense defeat most published prompt-injection defenses that reported near-zero success against static attacks. qualifies this findingResearch defenses with low static attack success failed under adaptive attack, so lab-reported rates against fixed attack sets may also overstate robustness.
- Measured hijack rates rise sharply when attackers get repeated attempts, so single-attempt figures understate risk. qualifies this findingWhere 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.
Status history
- 2025-05-20ReportedGoogle DeepMind: adversarial fine-tuning reduced but did not eliminate success. · record
- 2025-08-25CorroboratedAnthropic reports residual success after mitigations in its browser agent. · record
- 2026-02-05QualifiedMulti-attempt figures in Anthropic's system card are far higher than single-attempt rates. · record