| viXra.org > Artificial Intelligence > 2610 |
 |
 |
| viXra.org > Artificial Intelligence > 2610 |
|
|
[5] viXra:2610.0044 [pdf] submitted on 2026-10-10 12:44:19
Authors: Zohaib Muaz
Comments: 16 Pages. (Note by viXra Admin: Please submit article written with AI assistance to ai.viXra.org)
Autonomous Large Language Model (LLM) agents are increasingly vulnerable to indirect prompt injection (IPI) attacks, where adversarial instructions embedded in external data sources manipulate agent behavior without direct user input. These external data can originate from user documents, web retrieval results, or API outputs, leading to unauthorized compliance, sensitive information disclosure, identity spoofing, and cross-agent propagation of unsafe practices. This review synthesizes current research on IPI vulnerabilities and proposed mitigation strategies. Key attack vectors include manipulating webpage HTML via adversarial triggers for web agents, poisoning agent skills with tampered content, and injecting malicious commands into cloud logs processed by debugging agents. Covert attacks, which execute malicious actions without user-perceptible traces, pose a particularly insidious threat. Existing defenses, such as SecAlign's preference optimization, StruQ's structured queries, and multi layered zero-trust architectures, show promise. More advanced frameworks like ClawGuard enforce user-confirmed rule sets at tool-call boundaries, while ZEDD detects semantic shifts in embedding space for zero-shot detection. However, current guardrails often fail against sophisticated, adaptive attacks, and frameworks like AgentVigil and MUZZLE demonstrate high success rates in red-teaming exercises. The UReCoM attack further highlights vulnerabilities where benign users unknowingly relay adversarial content, bypassing existing defenses. The pervasive nature and increasing sophistication of IPI necessitate robust, behavior-grounded evaluation metrics and unified defense mechanisms to secure autonomous LLM agents in complex, untrusted environments.
Category: Artificial Intelligence
[4] viXra:2610.0041 [pdf] submitted on 2026-10-10 23:12:51
Authors: Tanmay Patil
Comments: 33 Pages. (Note by viXra Admin: Please submit article written with AI assistance to ai.viXra.org)
This paper presents an experimental protocol for testing whether an intermediate state in multi-digit addition has causal influence through hidden activations, visible reasoning tokens, both, or neither. Each arithmetic trace is generated by an executable interpreter and records the digit position, carry, partial result, next transition and final answer. The proposed test replaces a valid intermediate state and checks whether the remaining trace follows the corresponding counterfactual computation. Hidden-only, visible-token-only and joint interventions are compared with shuffled-donor, unrelated-donor and same-norm random controls at three positions around state-token emission. Evaluation uses full-trace counterfactual transition fidelity, nuisance invariance, repair, donor leakage and collateral change on separately held-out IID and out-ofdistribution splits. Attribution methods may identify candidate locations, but every retained location must pass an exact activation intervention. This paper reports the protocol and no new model results. Behavioural accuracy and a readable rationale alone cannot establish causal authority; the proposed interventions test it on held-out data.
Category: Artificial Intelligence
[3] viXra:2610.0040 [pdf] submitted on 2026-10-10 22:48:29
Authors: August Lau
Comments: 8 Pages.
AGI stands for artificial general intelligence and HSA stands for 3 areas (hypertopology/semigroup/acyclicity) in mathematics. The author will describe the connection between the two fields.
Category: Artificial Intelligence
[2] viXra:2610.0036 [pdf] submitted on 2026-10-10 00:43:25
Authors: Olegs Verhodubs
Comments: 7 Pages.
The bubble sort algorithm is one of the simplest and most wellu2011known in programming. That is why this algorithm was chosen as an experimental algorithm for studying the meanings embedded in it and their transformation. This research is part of a broader study whose final goal is to develop a system for the automatic generation of algorithms. Automatic algorithm generation can be implemented in various ways, but in this case, preference is given to a semantic approach — more precisely, an approach based on Semantic Web Technologies. The advantage of this approach lies in the direct, "willful" movement (from one concept to another) towards the result, as well as in the ability to trace the entire path from the initial data to the generated algorithm. To achieve the final goal, it is proposed to develop an ontology of the semantic subtext of the algorithm. This ontology will serve as a source of rules for an expert system for the automatic generation of algorithms.
Category: Artificial Intelligence
[1] viXra:2610.0013 [pdf] submitted on 2026-10-03 20:32:33
Authors: Clark M. Thomas
Comments: 3 Pages.
Questions swirling around human artificial intelligence, and visiting space-alien intelligence, are featured all over the media.Speculation clutters major idea gaps where scientific evidence, or at least high quality logic, should be. Ideas range fromoptimistic curiosity, to existential dread. Scientists can fairly well document our human past. Can we wisely envision and prepare for hybrid futures?
Category: Artificial Intelligence
| Third-Party Links: |
|