Adversa AI says its Coding Agent Security Platform for Developer Endpoints is generally available, and that it blocks dangerous coding agent actions before they run.
According to the October 29 release, the platform connects what coding agents and agentic IDEs such as Claude Code and Codex are asked to do with what they actually do in real time. Adversa AI says that link exposes attack chains and rogue agents whose individual steps look routine, and that the platform blocks the dangerous step before it executes.
The company lists four components. Observability and provenance use an agent behavior graph and a semantic threat layer. Policy enforcement and context guardrails cover real-time prevention. Runtime coverage spans an LLM proxy, an MCP gateway, agent hooks, the endpoint runtime and agent vendors’ compliance APIs. The fourth covers agentic supply chain security and shadow AI discovery for MCP servers, skills, hooks, plugins, instruction files and dependencies. Adversa AI says the engines evaluate each action against more than 300 default policies.
The release describes one deployment at a financial institution securing Codex, Claude Code and Copilot. Adversa AI reports a median blocking decision of 0.2 milliseconds, one confirmed incident per 2,000 raw signals, and a first-week block of an agent that loaded a malicious skill and tried to send out internal API secrets. These are the vendor’s figures and the release does not name the customer.
“Coding agents deliver enterprise AI’s quickest wins and biggest provable efficiency gains yet rank among the riskiest agentic AI systems according to AIRQ,” said Alex Polyakov, co-founder and CTO of Adversa AI.
In our reading, the useful detail for buyers is where the control sits. Coding agents run on developer laptops with access to credentials, so a control on the endpoint sees actions that a gateway in front of a model does not. Related coverage includes our opinion on giving every AI agent its own identity and our report on LiteLLM.
