Adversa AI launches platform to block cyberattacks on coding agents

Securing the AI developer

Adversa AI has launched a new security platform specifically designed to protect developer endpoints from cyberattacks involving coding agents. These agents, which include tools like Claude Code and Codex, are becoming central to how software is built. However, they also represent some of the riskiest AI systems currently in use.

The platform works by creating a link between what a coding agent is asked to do and what it actually does in real time. This visibility allows the system to identify 'rogue agents' that might appear to be performing routine tasks while actually executing a dangerous sequence of steps. By monitoring these actions, the platform can block a malicious step before it actually happens.

According to the company, the goal is to provide security teams with control while ensuring engineering teams maintain productivity. The platform aims to minimize the 'noise' of security alerts, which can often overwhelm IT departments. This balance is intended to help enterprises ship AI-driven software without introducing significant operational risks.

How the security engine works

The platform relies on an ensemble of in-memory engines to make rapid security decisions. These engines evaluate every action an agent takes against an agent behavior graph and a semantic threat layer. This architecture allows the system to check actions against over 300 default policies covering security, reliability, and privacy.

The technology is designed to detect attacks at a 0.1% false positive rate, which the company claims outperforms several major cloud and open-source guardrails. These benchmarks were conducted by organizations including Microsoft, ETH Zurich, and NVIDIA. The system is built to handle various agentic runtime surfaces, including LLM proxies, MCP gateways, and various compliance APIs.

The platform also addresses agentic supply chain security by identifying 'Shadow AI.' This involves detecting and managing all the various components an agent might use, such as MCP servers, skills, hooks, plugins, and instruction files. By mapping these into a universal behavior graph, the system can track every dependency and generated artifact produced by the AI.

Real-world performance results

A financial institution recently deployed the platform to secure tools like Copilot, Codex, and Claude Code. During this production test, the platform made blocking decisions in a median of 0.2 milliseconds. This speed was fast enough that the company reported no changes to the experience of the developers using the tools.

The security teams involved in the test were able to see every action taken by the agents from day one. The system's correlation capabilities were also noted for reducing noise, resulting in only one confirmed incident for every 2,000 raw signals. This level of filtering helps security professionals focus on genuine threats rather than false alarms.

During the first week of deployment, the platform successfully stopped a specific attack. In that instance, an agent loaded a malicious skill and followed hidden instructions to read internal API secrets. While other security tools missed the sequence of events, the platform blocked the final request to exfiltrate the data out of the system.

New threat categories identified

The defenses in the new platform are built on research conducted by the Adversa AI team into entirely new categories of AI attacks. These include techniques such as Reasoning Leakage, PromisqRoute, and Cryptographic Context Injection. The team has also identified threats they call SymJack and TrustFall, which are now being integrated into the broader industry threat landscape.

By focusing on these specific vulnerabilities, the platform addresses the unique risks posed by agentic AI. These risks differ from standard software bugs because they involve the way an AI reasons and interacts with its environment. The platform uses these insights to create context guardrails that prevent the execution of unauthorized attack chains.

Adversa AI has spent five years on offensive research to develop these protections. Their work has been recognized by various industry bodies and has appeared in major publications like The Wall Street Journal and Wired. The company continues to contribute to standards like the AARM standard by the CSA for agentic runtime security.