2026.04/AI Safety Theory
The Defense Trilemma: Why Prompt Injection Defense Wrappers Fail?
M. Bhatt, S. Munshi, V. S. Narajala, I. Habler, A. Al-Kahfah, K. Huang, J. Webb, B. Gatto, M. T. Hoque
Abstract
Proves that continuity, utility preservation, and complete safety cannot coexist for prompt-injection wrapper defenses, and characterizes the boundary where every such defense must fail.
Where it sits
Pillar 01 — Impossibility & Boundary Research
What can a defense never guarantee?
Prove where defenses cannot hold — what a control can guarantee, and which design directions are dead ends.
Outside writing
How outside writing is graded →More from limits
Whole pillar →2026.10The Coffee Law: Information, Reasoning and Communication Ceilings for Tool-Using Agent CollectivesAn AI safety result for multi-agent systems. Why do some agent collectives improve with more agents, more debate or a stronger model, while others do not? The productive question is which resource limits a collective, and which intervention removes that limit. The paper decomposes collective intelligence into three ceilings: information (a team cannot know more than its starting data plus what its tools can carry — talking and voting add nothing to that budget, and adding agents raises it only through more tool access), reasoning (what bounded agents can use is at most what they hold; debate and compute close the gap but never pass it, at a cost from linear to infeasible), and communication (evidence split across agents is unusable until enough is exchanged — about n/2 bits for some n-bit tasks even with unlimited computation). The decomposition explains published findings that otherwise look contradictory, yields a diagnostic for which ceiling binds, and makes predictions a single capability axis does not. Proofs are short; every finite example is verified exactly in public code.2025.06Bhatt Conjectures: On Necessary-But-Not-Sufficient Benchmark Tautology for Human Like ReasoningExamines why benchmark success may be necessary yet insufficient evidence of human-like reasoning, and formalizes the gap between measured performance and the claimed capability.