Cedric Clyburn from IBM Technology opened a recent video with a question that cuts straight to the core of AI coding security: if your developer used an AI tool to write a module, who supplied that code? Not the developer — not really. The AI did. And unlike every other software supplier in your vendor inventory, that supplier came with no contract, no SLA, no audit rights, and no software bill of materials.
That framing — AI as an unvetted software supplier — is the most useful mental model I've seen for thinking about what AI coding tools actually change about supply chain security. The video is worth watching. But 2026 has added a layer the video couldn't fully anticipate: attackers have already operationalized this attack surface, and they're doing it inside the tools themselves.
AI as a New Kind of Supplier
Clyburn's argument is structural. Traditional software supply chain security is built around the assumption that code comes from identifiable suppliers — developers, vendors, open-source maintainers — whose contributions can be tracked, signed, and verified. SBOMs (software bills of materials) work because you can enumerate what's in a software artifact and trace it back to a source.AI-generated code breaks that assumption. When a developer accepts a Copilot suggestion or scaffolds a service with an agentic coding tool, the resulting code has no clear provenance. It may be a synthesis of training data, documentation, Stack Overflow snippets, and patterns from millions of repositories. That synthesis isn't auditable in the traditional sense. There's no manifest, no signature chain, and no originating repo to verify against.
The IBM video identifies three things that change when AI enters the supply chain: the origin of the code itself, the dependency selection (AI tools suggest packages — including packages that may not exist), and the verification gap between what gets generated and what gets reviewed.
What 2026 Actually Looks Like
The attack community has not waited for defenders to catch up. The first half of 2026 produced a concentrated wave of supply chain operations specifically targeting AI coding infrastructure.The CSA documented five distinct campaigns between February and April: TeamPCP, ContextCrush, prt-scan, GlassWorm, and a compromise of the Bitwarden CLI package. The most significant, Miasma — attributed to TeamPCP — deployed a worm that disabled 73 Microsoft GitHub repositories before containment. It didn't target production systems directly. It targeted developer tooling and CI/CD pipelines, corrupting the build artifacts that flow into production downstream.
Mandiant's case study on Shai-Hulud is the other incident worth understanding in detail. The worm hijacked an active AI coding assistant session — not a dormant credential, an active session — and used it to traverse internal code repositories. The vector wasn't a misconfigured server or a leaked token. It was the AI coding tool itself, mid-session, with the developer present but not watching every output.
Intel 471 published analysis in the same window noting that attackers have shifted their primary targeting from package registries to IDE configuration files. Cursor and Claude Code were specifically named. The logic is straightforward: IDE config files contain credentials, endpoint mappings, and sometimes inline prompts that instruct the AI tool on how to operate in that environment. Compromising the config file compromises every session that runs against it.
The Hallucinated Package Problem
Separate from the active attack campaigns is a structural vulnerability that doesn't require an attacker to compromise any infrastructure at all. AI coding tools hallucinate package names. They suggest imports and dependencies with high confidence — including packages that do not exist in any registry.Researchers have named this slopsquatting: the practice of registering packages that match the hallucinated names AI tools commonly suggest. When a developer (or an agentic tool operating autonomously) installs a dependency based on an AI suggestion, they may be pulling from a package a threat actor seeded specifically because the name matched a known hallucination pattern.
A Black Duck survey in 2026 found that 95% of development teams are using AI-generated code, while fewer than 25% apply comprehensive IP, licensing, security, and quality checks to what they accept. That gap — near-universal adoption, low verification rate — is the structural condition that makes slopsquatting viable at scale.
Checkmarx data puts the vulnerability differential at 1.7 times more vulnerabilities in AI-generated code than in human-written code. Combined with autonomous agentic workflows where no human reviews individual pull requests before merge, that differential compounds. The DORA 2025 report found that 31% of pull requests across surveyed organizations received no human review at all.
The SBOM and SLSA Response
The industry response to AI supply chain risk is converging on two complementary frameworks: SBOMs for transparency and SLSA (Supply-chain Levels for Software Artifacts) for provenance verification.SBOMs are not new. What's new is the pressure to generate them for AI-assisted development workflows, where the traditional question ("what's in this artifact?") now needs to include: what portions were AI-generated, against what model version, with what training data, and with what dependency suggestions accepted without human review?
The EU Cyber Resilience Act mandates SBOM requirements effective December 2027. For US-based organizations, that deadline may feel distant, but many healthcare technology vendors with European operations or partnerships will need to comply — which means their downstream customers in US healthcare will increasingly be dealing with SBOM requirements as a vendor evaluation criterion, even without a direct regulatory mandate.
SLSA provides the complementary provenance layer. Where an SBOM says "here's what's in the artifact," SLSA says "here's how it was built, by what process, and with what verification." SLSA levels (1 through 4) represent increasing degrees of provenance assurance — from basic documentation through hermetic, reproducible builds with cryptographic attestation. The IBM video calls for independent review of AI-generated code as a core supply chain control; SLSA Level 3 and above represent the technical infrastructure that makes that review meaningful rather than aspirational.
| Framework | What It Addresses | Healthcare Relevance |
|---|---|---|
| SBOM | Artifact transparency — what's in the software | Vendor evaluation, incident response scoping |
| SLSA | Provenance — how the software was built and verified | Build pipeline integrity for clinical systems |
| Sigstore/Cosign | Artifact signing — cryptographic attestation of provenance | Container and package integrity verification |
| OWASP LLM Top 10 | AI-specific vulnerability taxonomy | AI coding tool risk classification |
What This Means for Healthcare
Healthcare development environments have specific exposure that the general supply chain security conversation doesn't always surface.HIPAA and AI-Generated Code
The HIPAA Security Rule doesn't distinguish between human-written and AI-generated code. If an AI coding tool generates a module that handles PHI, every requirement that applies to that module — access controls, audit logging, encryption in transit and at rest — still applies. What changes is the accountability structure: developers who accept AI suggestions without review are still accountable for the security properties of what they deploy. The "the AI suggested it" defense has no standing under HIPAA enforcement.Healthcare security teams should be working now to define what the SDL requires for AI-generated code that touches PHI-adjacent systems. That includes: does agentic coding that modifies a system in scope for HIPAA require a human review gate before merge? The answer should be yes. The question is whether that requirement exists in writing anywhere today.
Vendor Risk and SBOM Adoption
EU CRA timelines are a forcing function for the vendor ecosystem, even for US-only healthcare organizations. Any EHR vendor, medical device manufacturer, or clinical software vendor that has European customers will need to produce SBOMs meeting CRA requirements by December 2027. Healthcare procurement teams should be building SBOM requirements into vendor RFPs now — not because HIPAA requires it yet, but because the vendors who can't produce one in 2027 are the same vendors who haven't been tracking what's in their software today.The Miasma worm incident illustrates why this matters: 73 repositories compromised, all within a connected developer tooling ecosystem. If even one of those repositories feeds a clinical system, the SBOM gap becomes an incident response problem rather than a compliance planning problem.
IDE Security as a New Attack Surface
The Intel 471 finding about IDE config files deserves specific attention in healthcare environments. AI coding tools in clinical IT environments may be configured with endpoint mappings, model routing configurations, and prompt templates that touch the security architecture of the systems they're developing against. Those config files are now primary threat targets.Healthcare security teams that have done the work of evaluating AI coding tools for data handling and routing should extend that evaluation to include: where do the config files live, who has access to them, are they in source control, and have they been reviewed for inadvertent credential or architecture disclosure? The answer to at least some of those questions is probably unflattering.
The Verification Gap and Staffing
The bottleneck migration thesis — that AI-accelerated code generation is shifting the constraint to code review and security verification — lands especially hard in healthcare environments that are already short on AppSec resources. If 31% of PRs receive no human review in general engineering organizations, the number in healthcare IT environments, where AppSec bench depth is thinner and AI tool adoption is newer, is likely higher.Healthcare security teams building AI coding governance programs need to address the verification gap directly: who reviews AI-generated code, at what threshold, with what tooling, and under what documentation requirement. Without explicit policy on this, the default is no review — which is what the DORA data already shows at scale.
The Deeper Problem Is Verification
Clyburn closes the IBM video by calling for independent verification of AI-generated code — the same point ProjectDiscovery's 2026 AppSec survey captured in a single line: "Generation got cheap. Understanding got expensive."That framing connects directly to the structural argument in the Bottleneck post: code generation has become fast and cheap; the constraint has migrated downstream to review, verification, and accountability. AI supply chain security is where that constraint shows up concretely. SBOMs tell you what's in the artifact. SLSA tells you how it was built. Neither tells you whether the AI-generated logic is correct, safe, or aligned with the security properties the system requires. That's the gap that remains after the frameworks do their work — and it's a gap that requires human judgment at a moment when the tools creating the work are also creating pressure to skip the check.
Healthcare organizations that treat AI coding tools as productivity multipliers without building corresponding verification infrastructure will find that the productivity gain was borrowed against a security debt that doesn't announce itself until an incident.
This is an entry in the AI Security series. For related coverage, see The Bottleneck Just Moved to Security. Is Healthcare Ready?, When Your IDE Extension Is the Attack, and DPRK's AI Pivot: What Supply Chain Tradecraft Looks Like Now.
Key Links
- IBM Technology: Software Supply Chain Security — What AI Coding Changes (Cedric Clyburn) [Free]
- CSA: Miasma Worm Analysis — TeamPCP Campaign [Free]
- Intel 471: IDE Config Files as Primary Supply Chain Target [Free]
- Checkmarx: AI-Generated Code Produces 1.7x More Vulnerabilities [Free]
- Black Duck: State of Open Source and AI Code Adoption 2026 [Free]
- SLSA: Supply-chain Levels for Software Artifacts Framework [Free]
- EU Cyber Resilience Act — SBOM Requirements (December 2027) [Free]