Agentic AI for Autonomous Defense in Software Supply Chain Security: Beyond Provenance to Vulnerability Mitigation

1 st Toqeer Ali Syed, 2 nd Mohammad Riyaz Belgaum, 3 rd Salman Jan, 4 th Asadullah Abdullah Khan, 5 th Saad Said Alqahtani · 2025

Read Agentic AI for Autonomous Defense in Software Supply Chain Security: Beyond Provenance to Vulnerability Mitigation by 1 st Toqeer Ali Syed, 2 nd Mohammad Riyaz Belgaum, 3 rd Salman Jan, 4 th Asadullah Abdullah Khan, 5 th Saad Said Alqahtani (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-23480.txt.

Introduction

Toqeer Ali Syed and four co-authors argue that software supply chains need active defense, not just record keeping. Tools such as SLSA and software bills of materials prove where code came from, but they cannot find or remove vulnerabilities during production. Attacks increasingly target the trusted build and delivery steps themselves.

Their framework combines language-model agents that reason about code, reinforcement learning that balances security against build delays, and a blockchain ledger that records every action for auditing. The agents plug into real pipelines on GitHub Actions and Jenkins through a standard protocol. In tests covering injection attacks, unsafe deserialization, and access-control faults, the system detected more problems and mitigated them faster than rule-based or provenance-only baselines.

Catalog record

CREATOR1 st Toqeer Ali Syed, 2 nd Mohammad Riyaz Belgaum, 3 rd Salman Jan, 4 th Asadullah Abdullah Khan, 5 th Saad Said Alqahtani
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE32 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-23480

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Agentic AI for Autonomous Defense in Software Supply Chain Security: Beyond Provenance to Vulnerability Mitigation 1 st Toqeer Ali Syed, 2 nd Mohammad Riyaz Belgaum, 3 rd Salman Jan, 4 th Asadullah Abdullah Khan, 5 th Saad Said Alqahtani Abstract The software supply chain attacks are becoming more and more focused on trusted development and delivery procedures, so the conventional post-build integrity mechanisms cannot be used anymore. The available frameworks like SLSA, SBOM and in toto are majorly used to offer provenance and traceability but do not have the capabilities of actively identifying and removing vulnerabilities in software production. The current paper includes an example of agentic artificial intelligence (AI) based on autonomous software supply chain security that combines large language model (LLM)-based reasoning, reinforcement learning (RL), and multi-agent coordination. The suggested system utilizes specialized security agents coordinated with the help of LangChain and LangGraph, communicates with actual CI/CD environments with the Model Context Protocol (MCP), and documents all the observations and actions in a blockchain security ledger to ensure integrity and auditing. Reinforcement learning can be used to achieve adaptive mitigation strategies that consider the balance between security effectiveness and the operational overhead, and LLMs can be used to achieve semantic vulnerability analysis, as well as explainable decisions. This framework is tested based on simulated pipelines, as well as, actual world CI/CD integrations on GitHub Actions and Jenkins, including injection attacks, insecure deserialization, access control violations, and configuration errors. Experimental outcomes indicate better detection accuracy, shorter mitigation latency and reasonable build-time overhead than rule-based, provenance only and RL only baselines. These results show that agentic AI can facilitate the transition to self defending, proactive software supply chains rather than reactive verification ones. I Introduction The software supply chains are now one of the most important and risky but weakest links of the contemporary digital ecosystems. The recent and high-profile attacks like SolarWinds and the XZ Utils backdoor have shown that attackers can use trusted development and delivery flows to infringe thousands of downstream systems [ 19 ] . Such incidences demonstrate that security threats can no longer be limited to deployed software, but they are increasingly being introduced into the software production lifecycle, through an extended contribution of code, integration of dependencies, and pipeline build and deployment system [ 15 , 14 ] . Provenance, attestation, and transparency models such as the Supply chain Levels for Software Artifacts (SLSA), Software Bill of Materials (SBOM), and in toto offer fundamental underpinnings in terms of providing provenance and traceability of artifacts. Nonetheless, these mechanisms are mainly concerned with ensuring verification of post-production and a few abilities to respond in real-time or proactively prevent. When an illogical dependency, injected code, or, incorrectly configured component into the build pipeline, current tools are relatively reactive and usually cannot perform any interventions before artifacts have been damaged and distributed. The restriction puts emphasis on the necessity of a new layer of autonomous and intelligent defense systems that can learn, adapt, and operate during the process of software development. The agentic artificial intelligence (AI), comprised of self directed, goal oriented agents is one of the promising paradigms in this direction. Contrary to fixed rule based automation, agentic AI makes use of reasoning, self-reflection, and reinforcement learning (RL) to discover and reduce developing risks on changing environments. Within the software supply chain context, these agents have the capability to continuously monitor build environments, evaluate code provenance, evaluate dependency behaviour, and autonomously isolate or fix anomalies before release [ 31 ] . We present in this work an Agentic AI based Autonomous Defense Framework that builds upon the current supply chain security strategies beyond provingance to active vulnerability protection. The framework presents reinforcement learning empowered agents which persistently observe and mitigate runtime, configuration level threats, such as injection attacks, deserialization insecurity, broken access controls, system misconfigurations, in continuous integration and continuous deployment (CI/CD) processes. The proposed system provides adaptive learning, collaborative reasoning and explainable decision making by incorporating these agents via a shared security ledger which creates a closed feedback loop. The present work is one step towards the vision of self-defending software ecosystems supply chains which do not just record the integrity, but actively maintain it in real-time. II Background II-A Software Supply Chain and Its Vulnerabilities A software supply chain involves all the entities, processes, requirements that are associated with the development, building, trials, similarities, wrapping as well as deploying software systems. These are source code overlay, dependency overlay, continuous integration and continuous deployment (CI/CD) pipelines, artifact overlay, and update delivery overlay [ 14 , 15 ] . Every element presents a possible point of attack. The common pitfalls are caused by malicious dependency injection, hacked build servers, maladapted settings, or unauthorized commits on the code [ 16 , 17 ] . The 2020 attacks on SolarWinds backdoor and the XZ Utils hack have shown that attackers are no longer just depending on finding vulnerabilities in deployed software, but they are attacking development and build phases as well and corrupting trusted components even before the software is even deployed [ 18 , 19 ] . After being inserted into the pipeline, these attacks can spread downstream with signed, apparently legitimate artifacts, making them especially difficult to detect Once introduced into the pipeline, such attacks can spread downstream with signed, seemingly legitimate artifacts, which makes them especially hard to detect Introduced into the pipeline, such attacks can propagate downstream, containing signed artifacts that can appear legitimate, thus making their detection exceptionally difficult. [ 20 ] Introduced into the pipeline, such attacks can be spread, carrying signed artifacts that can look legitimate, Traditional mitigation approaches—such as code reviews, static analysis, or artifact signing—address integrity and traceability but remain largely reactive [ 21 ] . Frameworks like SLSA and in-toto strengthen provenance verification, yet they cannot autonomously prevent or respond to dynamic threats such as injection attacks, insecure deserialization, broken access control, or pipeline misconfigurations [ 22 ] . This gap motivates the need for adaptive, autonomous mechanisms capable of continuously learning and enforcing security across the development lifecycle [ 23 ] . II-B Agentic Artificial Intelligence The concept of agentic AI describes a system that consists of rational agents which sense their environment, think based on the surrounding contexts, and make actions outside towards specified security objectives [ 12 , 24 ] . In contrast to traditional automation scripts agentic system learns to change dynamically with changing environment via its closed feedback loop, that is, perception, cognition, action and learning (Franklin, 1997 autonomy). Each of these agents may act as a solitary actor or in groups, communicating with one another via communication protocols or knowledge ledgers shared between themselves [ 26 ] . The implementation of agentic AI in cybersecurity presents the possibility of self directed, explainable defense to cybersecurity vulnerabilities [ 34 , 35 ] . Repositories may be monitored, an analysis of commit patterns done, configuration may be verified, anomalies in build pipelines may be pointed out among other agents. Through the combination of cognitive reasoning and self reflection, they are able to justify their actions, reason the level of risk as well as cause changes in defensive policies in accordance with experienced learning. The given

Read the complete text