APEX: Active Protection at Execution Boundaries for LLM Agents

Evidence record

Research paper describing task-contract-based evidence gating and probe instrumentation against prompt injection at effect boundaries.

Read the original source ↗

Claim-level citations (3)

• supportsAPEX places prompt-injection checks at consequential execution boundaries: The APEX paper reports 0% attack success on five tested benchmarks and 0.56% on the sixth, under its configured test environment.

  Abstract; Table 3; effectiveness section

• supportsAPEX places prompt-injection checks at consequential execution boundaries: The proposed architecture compiles a trusted task contract before injection and uses WRAP and PLANT checks for external effects and unendorsed dependencies.

  Introduction and method: WRAP; PLANT; execution-boundary protection

• supportsAPEX places prompt-injection checks at consequential execution boundaries: The single reported SkillInject residual success involved attacker-authored guidance shaping the wording of an otherwise authorized economic report, a failure mode that did not create a separately identifiable prohibited external effect.

  Appendix SkillInject residual attack analysis: data storytelling SKILL.md IMPORTANT VALUE ALIGNMENT payload; 0.56% result