Findings/architectural-isolation-bounds-injection

Separating an agent’s control flow from untrusted data can give provable protection against specified forms of hijacking under explicit assumptions.

Qualifiedmeasured4 evidence records from 4 independent sourcesassistant-drafted
Scope: what this does not show

CaMeL’s protection is conditional on its trusted control flow and enforced policies; ToolFence is empirical and uses an LLM to grant capabilities. Multi-CaMeL shows that securing individual agents is insufficient when downstream delegation reinterprets untrusted data as instructions. Its benchmark tests control-flow integrity with no data-flow policies, and its implementation inherits incomplete implicit-dependency tracking. Utility costs and tested attacks depend on backend and environment; this is not a blanket injection-resistance guarantee. These are author-run benchmark studies, not independent deployment audits. ToolFence was tested against static injection strategies and retained non-zero failures. CapLite provides a separate, conditional write-integrity guarantee and reports 29.0% utility under attack versus 50.2% for its undefended baseline. The authors attribute most of the gap to planning by the small models. The utility cost depends on the task, model and control design.

Qualified: Still standing, but later work narrows how far it applies.

Evidence

Nov 12, 2026
CapLite reports blocked benchmark injections with a substantial utility trade-off
Xu, Liu and Nakayama / arXiv

Qualifies the utility boundary: CapLite reports no successful tool-goal attacks in 13,935 selected AgentDojo episodes, excluding output-only goals. Utility under attack is 29.0% on a broader task set. Its conditional guarantee covers request-grounded write integrity, not read confidentiality or every workflow.

How it relates to other findings

supportsqualifiescontestssupersedes
ReportedCorroboratedQualifiedContestedSupersededRevalidate· node size = evidence records · columns group by topic

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.

Key questions that rely on this finding

Status history

  1. 2025-03-24ReportedGoogle DeepMind introduces CaMeL. · record
  2. 2026-10-13QualifiedMulti-CaMeL narrows the single-agent guarantee: trust and provenance must be preserved across delegation, and the tested successor provides control-flow integrity under explicit assumptions rather than general data-flow protection. · record
  3. 2026-11-17QualifiedreconfirmedCapLite adds a narrower write-integrity guarantee with a substantial utility difference on its tested models, qualifying the earlier general wording about modest utility cost. · record