Results against fixed or known attacks, as reported by each defense's authors. StruQ's authors report 9% (TAP) and 58% (GCG) success on Llama for their strongest optimization attacks, and the instruction-hierarchy paper reports robustness gains rather than attack success rates. Says nothing about attackers who adapt to the defense. A September 2026 preprint (Speculative Safety Honeypot) adds 0.00% attack success on AgentDojo against fixed attacks only; its authors do not test attackers who know the defense is deployed. An October 2026 preprint finds an attention-based detector’s fixed-benchmark detection does not transfer to agent tool outputs.
Qualified: Still standing, but later work narrows how far it applies.
Evidence
The authors of a plug-in detection layer report 0.00% attack success on AgentDojo with a Qwen3-235B agent and two detectors, against 31.78% with no defense; the attacks were fixed benchmark attacks.
Qualifies the finding: Attention Tracker, an attention-based detector trained on synthetic data, averaged 0.97 AUROC on the deepset benchmark but 0.57 on a small AgentDojo subset of agent tool outputs (the authors’ AttnDA: 0.95 and 0.62), so benchmark detection results may not transfer to agent settings.
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 findingThe defenses' low static attack success does not carry over to adaptive attackers, which that finding's scope already excludes.
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
- 2024-02-09ReportedStruQ reports reduced injection success. · record
- 2024-03-20CorroboratedMicrosoft reports spotlighting cuts injection success below 2% on static attacks. · record
- 2025-05-20ContestedGoogle DeepMind finds adaptive attacks exceed 90% success against spotlighting-style defenses on Gemini. · record
- 2026-09-25QualifiedcorrectionAdaptive-attack results narrow this finding rather than dispute it, since its scope is limited to fixed attacks. The earlier entry's figure was wrong: spotlighting peaked at 82.4% under adaptive attack on Gemini, not above 90%. · record