AI adoption is having a profound impact on the software supply chain as AI tools have significantly reduced barriers to entry for programmers. However, AI-developed code is also introducing new security risks for organizations to tackle.

AI is now extensively used in the open-source ecosystem. A 2026 report by Stanford University counted 5.6 million open-source AI projects on GitHub and Hugging Face alone, triple the volume from 2023.

In addition, AI coding tools indiscriminately use open-source distributions, with models, frameworks and supporting services updating rapidly.

With around 85% of enterprise software today relying on open-source, it is certain that most, if not all, organizations are heavily reliant on AI-generated code.

It is vital that security teams understand the risks associated with AI-generated code and work with DevOps teams to integrate safeguards for modern AI workloads.

AI-Driven Software Development Risks

Vibe Coding

Vibe coding is the practice of using AI tools to generate code from prompts. It can significantly speed up the creation of software as well as lower the barrier to entry for developers.

However, the practice comes with significant security risks, from both malicious third parties and accidentally by developers themselves.

Attackers have developed techniques to exploit vibe coding. One such issue can occur when attackers register software packages with names hallucinated from AI systems.

This tactic is known as slopsquatting. It is a type of supply-chain threat in AI-powered workflows in which malicious actors can use the hallucinated, non-existent but plausible, package names to deliver malware.

Malicious payloads have been found to be hidden in AI models, container images and helper binaries, as well as being distributed through trusted platforms like Hugging Face.

Additionally, developers deploying legitimate code that has been produced using vibe coding can introduce security weaknesses.

For instance, AI models often generate code that is vulnerable to long-known security issues such as SQL injection and cross-site scripting (XSS).

Security misconfigurations are also common, with models sometimes recommending weak passwords, excessive privileges or overly permissive access controls.

Attackers Target AI Coding Tools

Researchers have observed significant efforts by malicious actors to target the AI coding tools now extensively used in software development.

Google research, published in September 2026, warned that the rapid integration of AI-assisted coding tools has become a primary target for threat actors and has contributed to several large-scale software supply chain compromises in 2025 and early 2026.

Attackers have developed a range of methods designed to insert malicious files into these tools, enabling them to deliver malware to organizations that use that code.

The Google researchers highlighted the activities of a financially motivated threat actor tracked as UNC6780 in this space. The group has conducted a series of large-scale open-source software supply chain compromises targeting ecosystems, including PyPI, npm and Docker Hub.

Attackers have also been observed exploiting flaws in AI coding tools to execute malicious code on victims’ machines. An example of this is a critical flaw discovered in open-source agentic AI coding tool OpenCode, which, if exploited, could allow an attacker-controlled remote model context protocol (MCP) server to execute arbitrary code on a developer’s machine during an MCP OAuth authentication process.

Overall, AI coding tools make it harder to have sufficient security checks and oversight over software deployed in organizations.

Austin Sedaghatpour, public sector mission lead at RapidFort, recently wrote: “Modern AI coding tools constantly change, use more open-source dependencies, and deploy more pervasively than traditional tools. Models, frameworks and supporting services update daily or weekly rather than quarterly or annually. Every non-hardened software artifact that is implemented by these tools magnifies the risk.”

Overwhelming Volume of Vulnerability Disclosures

The emergence of frontier AI models specialized in advanced cybersecurity, such as Anthropic’s Mythos and OpenAI’s GPT-5.4-Cyber, have significantly increased the volume of flaws identified.

This is particularly the case for open-source software, which is known to contain a particularly high volume of vulnerabilities.

Mike Wood, chief marketing officer, RapidFort, noted: “Frontier AI models can now scan a major open-source project and surface multiple vulnerabilities including chained impacts in a single pass within minutes.”

While advanced AI models are designed to benefit security teams, they also allow malicious actors to exploit flaws rapidly, often before organizations can apply fixes.

Disclosure-to-exploit windows have been compressed from 24 months a decade ago to less than 10 hours today.

Meanwhile, enterprise remediation cycles typically take weeks, with many vulnerabilities never addressed at all.

“This is not a theoretical future risk. It is the present condition of every system we are responsible for,” Wood said.

Consequences for Security Teams

The growing number of vulnerabilities introduced through the widespread use of AI in software development is having a significant operational impact on organizations.

RapidFort’s Sedaghatpour identified the following impacts that organizations are experiencing as a result:

• Restrictions on ATO-accessible technologies, with fewer AI tools approved for production
• AI deployment slowdowns as container stacks fail compliance on mission timelines
• Inability to compete with adversary AI tactics, increasing mission risks
• Recurring, high financial costs to manage vulnerability backlogs
• Highly skilled engineers focused on chasing and patching CVEs rather than delivering new capabilities

Software Security Must Move from Detection to Prevention

As AI brings significant changes the software supply chain attack landscape, new approaches to vulnerability management are required. A shift towards prevention rather than detection-based measures, such as software composition analysis (SCA), is vital.

Runtime intelligence and continuous monitoring have emerged as a means of making software development more transparent, auditable and trustworthy during live application.

As AI becomes more widely used in software development, modern production environments are becoming increasingly dynamic.

New CVEs are disclosed hourly, packages are updated, and processes change, while unauthorized modifications can occur after software has passed its early security checks.

Gaining runtime context enables teams to take a risk-based approach to vulnerability management that meets their organization’s specific needs. This is vital given the sheer volume of vulnerability disclosures today, which is beyond the human capabilities of security teams to manage.

For example, while a software package may contain dozens of CVEs, only a handful may affect the code paths used by an application. Some vulnerable components may never be executed, or even loaded into memory, during normal operation.

“Runtime context helps separate theoretical exposure from operational risk,” explained RapidFort’s Wood.

RapidFort Runtime Offerings

The RapidFort Runtime service offers a number of capabilities designed to support organizations develop a runtime intelligence strategy.

Continuous monitoring is a core capability of the RapidFort Runtime service. It identifies approved packages, binaries, libraries, processes and runtime behavior. This baseline becomes a reference point for pinpointing meaningful changes to software, such as a new package appearing inside a running workload or a previously unseen library being loaded.

RapidFort Runtime continuously observes the live production environment. Using these insights, it builds an evidence-based record of what is executing, identifies changes to deployed software, correlates new vulnerabilities with unique production workloads and provides the technical context needed to determine whether a newly disclosed risk actually matters.

A key aspect of this approach is the Runtime Bill of Materials (RBOM), which identifies the software components actually executing within production workloads using runtime instrumentation.

Organizations can use this RBOM information to understand real execution behavior rather than theoretical package inventories.

This differs from software bill of materials (SBOM), which lists everything included in an artifact. Static SBOMs also quickly become outdated.

The RF-Profiler tool generates and compares the RBOM with the SBOM for packages. It then suggests replacements from the curated library to reduce CVE counts and package size.

RapidFort Runtime also offers continuous CVE correlation, which evaluates newly disclosed vulnerabilities against software observed in an organization’s production environment.

This information can help security teams prioritize the remediation of vulnerabilities that pose an immediate risk to their organization’s live production environment.

Conclusion

In an era where AI generates code faster than security teams can assess it, the challenge is no longer finding vulnerabilities, but understanding which ones deserve immediate attention.

A proactive, preventative approach to software security is required. This can be achieved through continuous monitoring of attack surfaces and runtime intelligence, ensuring visibility of code during live application.

These insights enable security teams to quickly identify immediate risks to their production environments and prioritize their remediation strategies in a manner that meets their organization’s specific needs.