New tool watches AI coding helpers and blocks risky moves as they happen

Adversa AI, a security company based in Tel Aviv-Yafo, Israel, launched a tool on Oct. 29, 2026 that watches AI coding helpers and blocks a dangerous step before it runs. The company says the product is available now.

What it does

Coding agents are AI helpers that write and run software code for developers. Adversa says its platform compares what an agent was asked to do with what it actually does, in real time. Steps that look harmless alone can add up to an attack, and the tool links them together. The company names Claude Code and Codex as examples, and says one customer also used it with Copilot.

Adversa describes a case at a financial institution it did not name. In week one, it says, an agent loaded a malicious add-on, followed its hidden instructions, read internal API secrets (private keys that let software systems talk to each other) and tried to send them out. Adversa says other security tools missed it and its platform blocked the outgoing request.

It also reports that blocking decisions took a median of 0.2 milliseconds, a tiny fraction of a second, with no reported change for developers. It says its system cut noise to one confirmed incident per 2,000 raw signals.

What is not known yet

Every result here comes from Adversa's own announcement, and it describes only one customer. Pricing and independent testing were not shared. For any business using AI to write software, the practical question for the tech team is whether anyone is watching what those helpers do, and whether they can reach passwords or keys.