Research defenses and research models. CaMeL-style architectural isolation was not among the defenses tested in the main adaptive-attack study. A September 2026 preprint extends adaptive-attack results to 2026 API models (ASR@10 of 93.8% on GPT-5.6-Luna and 45.0% on GPT-5.6-Terra, and 7.3% on GPT-5.6-Sol by transfer without training on it), with no confidence intervals reported. An October 2026 preprint reports, from its authors’ own tests, an attention-gate defense that held adaptive-attack success to low but non-zero rates.
Reported: Stated by one source and not yet corroborated or challenged.
Evidence
Tested an undefended agent: novel attacks raised hijack rates from 11% to 81%. Shows attacks improve, not that defenses fail.
Adaptive attacks exceeded 90% against 2 of 8 baseline defenses on Gemini 2.0; a warning defense and a user-instruction classifier held (at most 11%). Shares authors with 'The Attacker Moves Second'.
Adaptive attacks exceeded 90% against most of 12 defenses; human red-teamers succeeded on every challenge in the subset of defenses they were given.
A different research group trains an adaptive attacker with curriculum reinforcement learning and reports 93.8% ASR@10 on GPT-5.6-Luna and 45.0% on GPT-5.6-Terra, and 7.3% on GPT-5.6-Sol by transfer without training on it.
Supports and qualifies the finding: in the Heads of Security preprint, RL-Hammer attackers reach 85% and 82% against SecAlign, while the authors report learned attention gates holding the highest success to 10% and 1% (from 97% and 99% undefended) on 100 held-out cases, measured at the first tool call and dependent on external trust labels.
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 qualifies Published prompt-injection defenses report attack success cut to near zero, or under 10%, against most of the fixed attacks their authors tested.The defenses' low static attack success does not carry over to adaptive attackers, which that finding's scope already excludes.
- This finding qualifies Frontier labs' layered defenses reduce prompt injection in tool-use, browser and computer-use agents but do not eliminate it.Research 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. supports this findingBoth show fixed, single-shot attack measurements understate attacker success; the attempt data comes from frontier models, not the published research defenses.
- Every frontier agent tested in two large public red-teaming competitions was hijacked at least once; in the 2026 competition the injections also had to conceal the compromise from the user. supports this findingCrowds of attackers who iterate get through every frontier agent tested; these competitions tested deployed models, not published research defenses.
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
- 2025-01-17ReportedUS AISI shows novel attacks raise hijack rates sharply. · record
- 2025-05-20CorroboratedGoogle DeepMind independently reports adaptive attacks above 90% on Gemini. · record
- 2026-09-25ReportedcorrectionThe 2025 US AISI and Google DeepMind entries did not test published defenses with near-zero reported success, and DeepMind shares authors with the primary study. 'The Attacker Moves Second' is the primary evidence; no independent replication is recorded yet. · record