XSci

Beyond the Patch: Reconstructing an Incomplete-Fix Patch Bypass in LLM Agent Frameworks with Control-Gated, Benign Probes

Chaitanya Vilas Garware

Submitted by Chaitanya Vilas Garware

Published November 19, 2026 · Version v1, November 19, 2026 · DOI 10.66977/xsci.2611.0034

Cybersecurity

Abstract

When a vulnerability at a security boundary of an LLM agent framework is patched, the advisory identifies the affected and patched versions, and replaying the proof-of-concept (PoC) tests only whether its reported trigger is still reproducible. That the trigger no longer fires does not, on its own, show that the underlying security invariant holds on every code path that reaches the boundary. We present a small, control-gated reconstruction method and a pilot study built around one question: can original-trigger replay report a boundary as fixed while the same invariant stays violable through a distinct, separately reachable path at the same version? We compare original trigger replay (B0) with a distinct sibling-path probe (M) for PraisonAI, and perform five additional before/after reconstructions across Langroid, Agno, and CrewAI. Each run emits a tri-state verdict of VIOLATION, NO_VIOLATION_OBSERVED, or INCONCLUSIVE. A mandatory pos itive (authorized) control gates the verdict, so that a non-violation result is admitted only when the control passes in the same run. All probes are benign, using an in-memory flag, a filesystem sentinel, or a builtins reachability check, and none performs a real payload. Our central result reproduces an officially recognized incomplete fix. In PraisonAI, original trigger replay reports the tool-loading boundary as fixed at version 4.6.31, yet the invariant is still violated through the tool_override path. That residual was later assigned its own identifier (CVE-2026-44334, catalogued as a “CVE-2026-40287 patch bypass”) and closed at 4.6.32. The five before/after measurements span four frameworks and four invariant classes. For a CVE whose advisory records no fixed version, we observe confinement in tested release 1.11.1 after observing a violation in 1.11.0. We are explicit about scope. A non-violation verdict means “no violation observed under the tested probe, version, and configuration,” not a proof of safety. The headline harness is control-gated per path, and fourteen fault-injection scenarios produced their expected verdicts. Only one same-version B0̸=M divergence is demonstrated. The PraisonAI probe invokes the documented loader functions directly, which establishes boundary behavior under the supplied setup rather than a full end-to-end exploit. Finally, we do not claim the method is a new testing technique or that it estimates prevalence.

View PDF