{"slug":"one-unverifiable-agent-breaks-the-whole-chain-proves-new-paper","citations":[{"url":"https://arxiv.org/abs/2603.14332","committed_hash":"sha256:2ad19165fc45fe9441baeab934f9bcbf596d0df6458605ad27fb9d8df154a217","committed_hash_short":"sha256:2ad19165…f154a217","mime_type":"text/html","committed_at":"2026-09-19T21:00:25.385316+00:00","content_snapshot":"\n\n\n [2603.14332] Governing Dynamic Capabilities: Cryptographic Binding and Reproducibility Verification for AI Agent Tool Use\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n \n \n \n \n \n\n\n
\n Skip to main content\n \n \n \n
\n \n\n \n \"archive\"\n \n\n \n\n \n
\n\n\n
\n
\n
\n\n
\n
\n

Computer Science > Cryptography and Security

\n
\n\n
\n arXiv:2603.14332 (cs)\n
\n\n
\n
\n
\n [Submitted on 15 Mar 2026 (v1), last revised 19 Mar 2026 (this version, v2)]
\n

Title:Governing Dynamic Capabilities: Cryptographic Binding and Reproducibility Verification for AI Agent Tool Use

\n
Authors:Ziling Zhou
\n View PDF\n HTML (experimental)\n\n\n\n
\n Abstract:AI agents dynamically acquire tools, orchestrate sub-agents, and transact across organizational boundaries, yet no existing security layer verifies what an agent can do, whether it executed what it claims, or what happened in a multi-agent interaction. We trace this gap to the capability-context separation: inside a transformer, tool definitions and user context are indistinguishable tokens, but at the orchestration layer they have fundamentally different security semantics. Existing frameworks conflate the two, enabling silent capability escalation and leaving interactions without verifiable provenance.\n
From this principle we derive three Agent Governance Requirements: capability integrity (G1), behavioral verifiability (G2), and interaction auditability (G3), defining what a governed agent ecosystem must enforce, independent of how. We prove two structural results: the Chain Verifiability Theorem (one unverifiable interior agent breaks end-to-end verification for all downstream nodes) and the Bounded Divergence Theorem (replay-based verification yields a probabilistic safety certificate, epsilon <= 1 - alpha^{1/n}). We validate with two crypto-agnostic instantiations -- basic (Ed25519, SHA-256; 97 us verify) and enhanced (BBS+ selective disclosure, Groth16 DV-SNARK; 13.8 ms) -- both satisfying nine security properties. A reproducibility study (9 models, 7 providers) reveals 5.8x variance in inference determinism, connecting model characteristics to governance architecture. End-to-end evaluation over 5-20 agent pipelines confirms <0.02% overhead and detection of all attack scenarios with zero false positives.\n
\n\n \n
\n \n \n \n \n\n \n \n \n \n \n \n \n \n \n \n \n \n \n
Comments:23 pages, 5 figures, 18 tables. Includes 11 experiments, 9 formal security properties, and appendix with proofs
Subjects:\n Cryptography and Security (cs.CR)
Cite as:arXiv:2603.14332 [cs.CR]
 (or \n arXiv:2603.14332v2 [cs.CR] for this version)\n
  https://doi.org/10.48550/arXiv.2603.14332
\n \n \n
\n arXiv-issued DOI via DataCite
\n
\n
\n
\n
\n
\n
\n

Submission history

From: Ziling Zhou [view email]
[v1]\n Sun, 15 Mar 2026 11:46:57 UTC (311 KB)
\n [v2]\n Thu, 19 Mar 2026 19:46:56 UTC (323 KB)
\n
\n
\n \n
\n \n Full-text links:\n

Access Paper:

\n
    \n
  • View PDF
  • HTML (experimental)
  • TeX Source\n
\n
view license
\n
\n
\n

Current browse context:

\n
cs.CR
\n\n
\n\n < prev\n \n   |   \n next >\n
\n
\n new\n | \n recent\n | 2026-03\n
\n Change to browse by:\n
\n cs
\n
\n
\n\n
\n
\n

References & Citations

\n \n
\n
\n\n
\n \n \n
\n\n
\n

Bookmark

\n \"BibSonomy\"/\n \n \n \"Reddit\"/\n \n
\n \n\n
\n
\n \n
\n

Bibliographic and Citation Tools

\n
\n
\n
\n \n
\n
\n Bibliographic Explorer (What is the Explorer?)\n
\n
\n
\n
\n \n
\n
\n Connected Papers (What is Connected Papers?)\n
\n
\n
\n \n
\n
\n Litmaps (What is Litmaps?)\n
\n
\n
\n
\n \n
\n
\n scite Smart Citations (What are Smart Citations?)\n
\n
\n
\n
\n
\n
\n
\n
\n\n\n \n \n
\n

Code, Data and Media Associated with this Article

\n
\n
\n
\n \n
\n
\n alphaXiv (What is alphaXiv?)\n
\n
\n\n
\n
\n \n
\n
\n CatalyzeX Code Finder for Papers (What is CatalyzeX?)\n
\n
\n\n
\n
\n \n
\n
\n DagsHub (What is DagsHub?)\n
\n
\n \n
\n
\n \n
\n
\n Gotit.pub (What is GotitPub?)\n
\n
\n\n
\n
\n \n
\n
\n Hugging Face (What is Huggingface?)\n
\n
\n\n
\n
\n \n
\n
\n ScienceCast (What is ScienceCast?)\n
\n
\n
\n\n
\n
\n
\n
\n
\n
\n
\n\n\n \n \n
\n

Demos

\n
\n
\n
\n \n
\n
\n Replicate (What is Replicate?)\n
\n
\n
\n
\n \n
\n
\n Hugging Face Spaces (What is Spaces?)\n
\n
\n
\n
\n \n
\n
\n TXYZ.AI (What is TXYZ.AI?)\n
\n
\n
\n
\n
\n
\n
\n \n \n
\n

Recommenders and Search Tools

\n
\n
\n
\n \n
\n
\n Influence Flower (What are Influence Flowers?)\n
\n
\n
\n
\n \n
\n
\n CORE Recommender (What is CORE?)\n
\n
\n
\n
\n \n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n\n \n \n
\n
\n
\n

arXivLabs: experimental projects with community collaborators

\n

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

\n

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

\n

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

\n
\n
\n

\n
\n
\n
\n\n
\n
\n\n
\n Which authors of this paper are endorsers? |\n Disable MathJax (What is MathJax?)\n \n
\n \n
\n
\n
\n\n
\n\n \n\n\n\n","snapshot_chars":42705,"live_check":"matches"}]}