See how Checkmarx can enhance your AI security at the speed of development Use AI SCS for continuous discovery and policy enforcement. Start by taking an inventory of all AI usage to uncover Shadow AI. AI agents introduce security risks including excessive privilege, insecure tool invocation, indirect prompt injection via external data sources, and unintended data exfiltration.
Put simply, AI supply chain security matters more now because AI systems are more connected, more exposed, and more consequential than they were even a short time ago. That means attackers don’t need source code or file access to interact with AI systems – they can probe them directly. Modern AI systems depend on pre-trained models, open-source libraries, training datasets, vector databases, and managed cloud services.
Discover our open-source tool designed to proactively identify vulnerabilities in MCP servers. AI Defense connects to enterprise Model Context Protocol (MCP) servers and repositories to proactively identify http://4dw.net/jqueen/privacy.php threats like tool poisoning, data exfiltration, and arbitrary code execution. Because AI agents are more capable and autonomous systems, they rely on a variety of protocols, tools, and resources. Instead of flagging individual misconfigurations, Wiz reveals attack paths – the combinations of exposure, permissions, and data access that turn small issues into real supply chain risk. Using an agentless approach, Wiz continuously discovers AI services, training environments, model endpoints, storage locations, and supporting infrastructure across AWS, Azure, and GCP.
Generate AI BoMs
Organizations should treat public registries as untrusted sources — scan all downloaded models with ModelScan, verify provenance, prefer safetensors format, and maintain a curated internal registry for production use. The primary defense is using ModelScan to detect malicious code and migrating to safetensors format, which cannot execute code during loading. Python’s pickle serialization can execute arbitrary code during deserialization, and attackers have demonstrated uploading trojanized models to public registries. The most common attack vector is malicious code embedded in pickle-format model files. Regular training sessions and tabletop exercises build the awareness needed to maintain a secure supply chain.
Workflow Integration
- This discovery effectively creates an AI Bill of Materials (AI BOM), helping teams inventory models, datasets, pipelines, identities, and services – and surface shadow AI that falls outside standard controls.
- US national cyber defense agency publishing AI security guidance and protecting critical infrastructure from AI threats.
- Many organizations discover they have 3-5x more AI deployments than initially estimated.
- Fickling is a specialized tool for analyzing and decompiling Python pickle files, allowing security teams to inspect the exact code that would execute when a pickle-based model file is loaded.
- A comprehensive guide to LLM security — vulnerabilities, the OWASP Top 10 for LLMs threat landscape, API security, supply chain risks, monitoring, and defense strategies for large language models.
Backdoors embedded in model weights that activate on specific inputs. Model cards can misrepresent training data, evaluation metrics, or intended use. Catalog size and download counts change continuously; the durable control is provenance for the exact artefact deployed. Data-lineage capabilities may support that work, but no single lineage tool or artifact establishes compliance.
- Without a unified view, teams may secure individual environments while missing cross-cloud attack paths.
- PyPI is a large public package ecosystem whose account, publishing, provenance, and malware defenses continue to evolve.
- AI assets are often created outside of traditional software workflows – by data science teams, experimentation notebooks, or ad-hoc pipelines.
- Assess third-party AI services with the same scrutiny as critical infrastructure.
- Connect a consequential decision point to existing dependency scanning, signature checking, model acceptance, or inventory controls.
- GLACIS does not generate a complete ML-BOM, replace a scanner, or establish upstream provenance by itself.
OWASP Top 10 for Large Language Model Applications
Custom operators, serialization formats, and GPU kernel code can introduce vulnerabilities that https://northfloridahouse.com/vpn-for-onlyfans-possibilities-and-advantages-of-use.html traditional scanners miss. The typical AI deployment depends on hundreds of software packages, each representing a potential attack vector. AI infrastructure introduces supply chain risks at the framework, library, and hardware levels. Adversaries can create web pages containing carefully crafted text designed to influence models that train on web crawls. The datasheets for datasets framework (Gebru et al., 2021) provides a standardized template for dataset documentation. Hash verification is the simplest provenance check — compare the cryptographic hash (SHA-256) of a downloaded model file against the hash published by its creator.
Model Provenance and Integrity
AI supply chain security isn’t a separate discipline from cloud security – but it also isn’t limited to the cloud alone. Without a unified view, teams may secure individual environments while missing cross-cloud attack paths. Finally, securing AI supply chains at scale is harder in multi-cloud environments.