{
 "license": "CC-BY-4.0",
 "attribution": "Fide AI, Agentic Cyber Explorer",
 "url": "https://agentic-cyber-explorer.pages.dev/events/caplite-request-grounded-write-integrity-2026/",
 "asOf": "2026-11-17",
 "id": "caplite-request-grounded-write-integrity-2026",
 "date": "2026-11-12",
 "datePrecision": "day",
 "title": "CapLite reports blocked benchmark injections with a substantial utility trade-off",
 "lane": "defense",
 "kind": "paper",
 "summary": "CapLite derives write permissions from tool schemas and values in the user request before reading untrusted tool results. The preprint’s authors report no successful tool-goal attacks in 13,935 AgentDojo episodes across 15 open-weight models, and 29.0% utility under attack on a broader task set. Its conditional write-integrity guarantee excludes output-only goals and does not establish read confidentiality or general agent security.",
 "whyItMatters": "It tests whether enforceable restrictions can protect smaller models while making the cost of restricted workflows visible.",
 "actors": [
  "university-of-tokyo",
  "institute-of-automation-cas"
 ],
 "topics": [
  "prompt-injection",
  "access-controls",
  "eval-validity"
 ],
 "atlas": [],
 "artifacts": [
  "caplite",
  "agentdojo",
  "camel"
 ],
 "sources": [
  {
   "url": "https://arxiv.org/abs/2611.17115v1",
   "publisher": "Xu, Liu and Nakayama / arXiv",
   "title": "CapLite: Prompt-Injection Defense for Small Language Model Agents without Model-Written Security Artifacts",
   "date": "2026-11-12",
   "type": "primary",
   "accessed": "2026-11-15"
  }
 ],
 "keyFacts": [
  {
   "fact": "The evaluation covers 15 open-weight models, 13 between 1.7 and 12 billion parameters, using AgentDojo tasks and selected attacks.",
   "locator": "Abstract; experimental setup"
  },
  {
   "fact": "Under the main published attack, no tool-goal attack succeeded across 15 × 929 episodes (13,935 total). Attack success excludes 20 output-only pairs per model, which CapLite does not protect. This is not universal adaptive-attack resistance.",
   "locator": "Section 4: benchmark; Table 1"
  },
  {
   "fact": "Utility under attack averaged 29.0%, versus CaMeL’s 27.2% and the undefended agent’s 50.2%, on all 949 pairs per model, including the output-only pairs excluded from attack-success scoring. The authors attribute most of the gap to planning by the small models.",
   "locator": "Section 4; Tables 1–2; utility results; Appendix D"
  },
  {
   "fact": "CapLite commits write calls before reading tool results and permits security-relevant arguments only when grounded in the user request. Its proof is conditional on this enforcement model and addresses write integrity, not unrestricted workflows or read confidentiality.",
   "locator": "Write-integrity guarantee; limitations"
  }
 ],
 "significance": 4,
 "fideQuestions": [
  "FID-075"
 ],
 "methods": [
  "capability-restriction",
  "control-data-isolation"
 ],
 "review": "assistant-drafted",
 "addedOn": "2026-11-15",
 "revisions": [
  {
   "id": "clarify-utility-gap",
   "on": "2026-11-17",
   "kind": "correction",
   "why": "Correction: the utility comparison measures the complete tested configurations. The authors attribute most of the observed gap to small-model planning."
  }
 ]
}