| Spotlighting / datamarking | Delimit, mark, or encode untrusted text | Among the weakest adaptively; "barely stopped attacks"; ~28–42% residual ASR |
| Sandwich / instruction hierarchy | Repeat rules around untrusted content; rank user above data | 500-round study: every model-reliant layer fell (hierarchy 247, sandwich 277); abrupt breaks, so short evals overstate |
| Injection classifiers / guard models | Detector flags known injection patterns | GCG universal suffix fools probes 94–99%+; 10% FPR still allowed >90% ASR |
| LLM sanitizer (strip injection) | Second LLM removes injected text first | Strong on one benchmark (54% to 0%) but the guardrail LLM is itself injectable; weak backends degrade below no-defense |
| Tool-boundary firewalls | Minimize args / sanitize tool I/O at the call | Rare encodings slip (Braille beat a sanitizer); benign utility drops notably |
| Dual-LLM / quarantine | Privileged planner never sees untrusted data | Quarantined LLM still injectable; can encode exfil into its own output |
| CaMeL (capability / data-flow) | Plan from trusted query only; interpreter tracks provenance; policies gate calls | Holds control-flow (300 to 0) but residual: in-band harmful actions, phishing in authorized replies, composition of allowed flows, ~1-bit side channel, ~2.7× token cost |
| Rule of Two (capability cap) | At most two of: untrusted input, sensitive access, external action | Not an input filter; forces a utility tradeoff; doesn't cover all agentic risk |
| Output filter / DLP | Block responses carrying secrets; block exfil channels | The one layer that never leaked in the 500-round study, but substring scoring misses encoded exfil (upper bound) |
| Adversarial training | Fine-tune on simulated injections | Non-adaptive drops look great (92% to 18%); adaptive attacks recover most |
| Confirmation / HITL gating | User approves consequential actions | Degrades through fatigue; high-frequency prompts train click-through |