JFrog expands development lifecycle security for agentic era

JFrog is expanding its software supply chain security capabilities to address the risks associated with AI-generated code and autonomous AI agents, as businesses face growing pressure to maintain security, governance and regulatory compliance while accelerating software development.

The company has introduced capabilities designed to protect software development environments, automate vulnerability remediation and provide evidence of compliance throughout the software development lifecycle.

Sunny Rao, SVP, APAC at JFrog, said the announcements made at the company's SwampUp event earlier this year reflected a fundamental shift in how software is developed, with human-written code increasingly giving way to code generated by AI agents.

"At JFrog, we have created the software supply chain platform that protects, remediates, and controls itself," Rao said.

Rao said enterprises need to maintain trust in software produced through increasingly automated development processes, while ensuring that security controls and governance policies keep pace with the speed of AI.

Rao identified three priorities for organisations adopting agentic AI: protecting software environments from malicious components, remediating vulnerabilities at machine speed, and governing the activities of AI agents throughout development and deployment.

The changes come as organisations seek to translate AI investment into productivity gains without introducing additional security and compliance risks.

Establishing single source of truth

Visibility across the software development lifecycle is becoming increasingly important with agents capable of generating code, accessing external resources and performing tasks with varying degrees of autonomy.

Rao argued that organisations need a central record of agent activity, including the instructions provided by humans, the resources accessed, the code produced and the approvals required before software could reach production.

Without this visibility, organisations could struggle to establish whether an AI-generated component complied with internal policies or determine how a particular change progressed through development, testing and deployment.

"When you get into agentic development, the single most important aspect, and something that the industry is grappling with, is how do you trust what comes out of it?" Rao said.

A centralised system of record could provide the evidence required to establish that software had been developed and deployed according to an organisation's policies.

This includes understanding what an agent was instructed to do, which permissions it received, what resources it accessed and whether any exceptions to established procedures were authorised.

This information is particularly important in the context of changing regulatory frameworks, with organisations required to demonstrate compliance more quickly.

Previously, compliance teams might have taken weeks to assemble supporting evidence from spreadsheets, emails, collaboration platforms and project management systems.

That approach is becoming impractical as regulators increasingly expect organisations to provide evidence within hours.

Agents introduce new security risks

Organisations need to establish appropriate guardrails around AI agents rather than treating unexpected agent behaviour as an isolated technical problem, Rao said.

He argued that AI agents generally acted according to the objectives, instructions and permissions established by humans, placing responsibility on organisations to ensure those systems operated within appropriate boundaries.

"Accountability still sits with the human," Rao said.

He said software developers are increasingly responsible for creating not only code but also the environments, policies and security controls within which AI agents operate.

One emerging concern involves the use of agent skills, which are collections of instructions and executable commands that extend an AI agent's capabilities.

Developers may download these skills from public repositories or marketplaces to automate tasks, but malicious instructions can potentially be concealed within otherwise legitimate-looking content.

Rao cited one example involving a skill containing thousands of lines of padding, with a malicious command concealed deep within the material.

Such a command could instruct an agent to retrieve and execute content from the public internet within an environment where it had access to sensitive resources.

The example showcases how ostensibly harmless extensions can create security risks when their instructions are not adequately inspected.

Rao also highlighted the importance of controlling how agents connect to enterprise systems through MCP gateways.

These connections can give agents access to tools and organisational resources, making it important to ensure their permissions are limited to what is necessary for the task.

JFrog's approach is intended to help organisations enforce these policies across their software development environments, including the components and plugins used by both human developers and AI agents.

Its objective is to limit the ability of agents to act outside their intended scope while allowing enterprises to take advantage of automated development.

Zero-touch remediation targets quicker vulnerability fixes

JFrog is also targeting the growing volume of security issues that can emerge as AI accelerates software production.

Rao said organisations had initially expected automated code generation to deliver substantial productivity improvements, but those gains could be undermined by the additional work required to review and secure the resulting software.

He said code generation had become significantly faster, while human-led code reviews and security assessments struggled to keep pace.

The resulting volume of code can create code fatigue, with reviewers struggling to assess every change thoroughly.

Security teams then face another challenge when testing identifies large numbers of vulnerabilities, including false positives and issues that may not be exploitable within a particular application or environment.

JFrog's zero-touch remediation is designed to help organisations prioritise and address vulnerabilities with greater automation.

The process combines protection at the point where software and components enter an environment, with continuous analysis throughout development.

Contextual analysis can then help determine which vulnerabilities are relevant to a particular application, which are exploitable and which pose the greatest risk.

Zero-touch remediation is intended to automate the deployment of available fixes through the affected environment, with the aim of addressing high-priority vulnerabilities within minutes.

That reduction in response time is becoming increasingly important because attackers are now using advanced AI models to identify vulnerabilities and develop exploits.

"The ability to take what has been a fix that has been released in the public repositories, checking that and making sure that it propagates right through your environment, especially where the highest risk is, and address it in minutes becomes very important," Rao said.

Compliance evidence built into development

The third component of JFrog's strategy focuses on governance, risk and compliance, areas that are becoming increasingly important as organisations adopt agentic development.

These teams are emerging as key stakeholders because they need to establish an organisation's security posture and demonstrate compliance with applicable requirements.

Rao highlighted the evolving Australian regulatory environment as another consideration for businesses deploying AI and software-intensive systems.

For multinational organisations, compliance can become more complicated because different jurisdictions may impose different requirements.

JFrog's approach is to collect evidence as software moves through its development lifecycle rather than requiring organisations to reconstruct events after a problem occurs.