JFrog offers a feature called Local Shadow AI, designed to detect on developers’ workstations the AI tools, skills and components that bypass corporate approval processes. In an interview with Monde Informatique, Nir Peleg, global director of information security for field teams at JFrog, says that the proliferation of coding agents is creating new risks for the software development chain.
Coding agents generate code directly from prompts and can download packages or skills themselves whose reliability has not been verified by the company. These components may contain vulnerabilities or malicious instructions. According to Nir Peleg, the feature scans development environments to identify unauthorized items and the risks they pose. Security teams can then assess the detected threats and define the necessary remedial actions.
Behavior‑based analysis
JFrog says it has developed its own analysis mechanism, based on multiple methods. It can spot, for example, a skill that attempts to obtain privileges or accesses it does not need to perform its function. The danger also lies in instructions hidden behind sophisticated language: a skill may appear legitimate while steering the agent’s behavior toward undesirable actions. The company therefore aims to detect behavioral anomalies, and its research teams update the methods as techniques evolve.
Some customers already have a large language model (LLM), i.e., an AI model trained on vast volumes of text, approved by their organization. The Bring Your Own LLM Key (BYOK) approach allows them to use their own API key, which provides access to that model, rather than necessarily relying on JFrog’s service. The analysis mechanism then leverages this model, giving companies the freedom to choose their provider and retain their security agreements and policies. In a self‑hosted deployment, data remains in the client’s environment and code is not sent to JFrog’s cloud.
Autonomous attacks and governance
Among the scenarios that worry Nir Peleg is the ability of cutting‑edge models to contribute to autonomous attacks. They can help discover new vulnerabilities, including zero‑day types, i.e., flaws still unknown to the vendor, and then participate in multi‑step attack chains. “Attackers can potentially operate at an unprecedented scale and speed,” while defenders still have to deal with patch, validation and approval cycles that can take weeks.
To govern these uses, the first step is to gain visibility into what is happening. “If you don’t see Shadow AI downloads or the components used by agents, you can’t govern them,” he says. Then security and control policies must be integrated into the development process, whether it concerns dependencies, skills, MCP connectors (protocol linking an agent to external tools) or deployments.
The goal is to have a common system that centralizes visibility, security analyses, rules and controls, in order to trace every step from code modification to production. When development occurs at the speed of agents, governance must keep pace: it can no longer rely solely on post‑hoc human checks, and controls must be embedded in the platform and automated processes.
