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.
It tests whether enforceable restrictions can protect smaller models while making the cost of restricted workflows visible.
Key facts
As stated in the sources, with where to find them.
- The evaluation covers 15 open-weight models, 13 between 1.7 and 12 billion parameters, using AgentDojo tasks and selected attacks.Abstract; experimental setup
- 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.Section 4: benchmark; Table 1
- 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.Section 4; Tables 1–2; utility results; Appendix D
- 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.Write-integrity guarantee; limitations
Record updates
Correction: the utility comparison measures the complete tested configurations. The authors attribute most of the observed gap to small-model planning.
Findings that cite this record
Key questions this bears on
- 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.