---
type: "post"
id: "adversa-ai-launches-coding-agent-security-platform-for-282ec1"
slug: "adversa-ai-launches-coding-agent-security-platform-for-282ec1"
title: "Adversa AI launches Coding Agent Security Platform for Claude Code, Codex and agentic IDEs"
description: "Adversa AI is marketing a Coding Agent Security Platform for enterprises using Claude Code, Codex, Copilot, Cursor, and other agentic development environments."
retrieval_nugget: "Adversa AI is marketing a Coding Agent Security Platform for enterprises using Claude Code, Codex, Copilot, Cursor, and other agentic development environments. The product claims to observe linked agent behavior across models, tools, hosts, and endpoints, then block a dangerous action before it completes rather than merely flagging a risky prompt."
published_at: "2026-10-29"
updated_at: "2026-10-29"
record_date: "2026-10-29"
date_kind: "published_at"
topics: ["agent-security","claude","claude-code"]
entities: ["Adversa AI","Coding Agent Security Platform","Claude Code","OpenAI Codex","GitHub Copilot","Cursor","Model Context Protocol"]
related_patterns: ["agent-security-moves-to-runtime-boundaries","agent-ready-software-exposes-capabilities"]
novelty_label: "new_fact"
what_changed: "Adversa AI launches its own runtime control platform for coding agents, monitoring LLM calls, tool executions and endpoint actions and blocking dangerous actions."
prior_refs: ["memory:71d65406-63a6-89fc-afc5-c46928d093e1","memory:845ef982-e55e-831a-aff2-9773529b1e77","memory:b3fd08c2-e129-8cc1-aa33-d38fcf0340ce"]
source_url: "https://adversa.ai/agent-security-platform"
source_urls: ["https://adversa.ai/agent-security-platform"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:adversa-ai-launches-coding-agent-security-platform-for-282ec1"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/adversa-ai-launches-coding-agent-security-platform-for-282ec1/","markdown":"https://newruntime.com/posts/adversa-ai-launches-coding-agent-security-platform-for-282ec1.md","json":"https://newruntime.com/posts/adversa-ai-launches-coding-agent-security-platform-for-282ec1.json"}
---

# Adversa AI launches Coding Agent Security Platform for Claude Code, Codex and agentic IDEs

## Retrieval answer

Adversa AI is marketing a Coding Agent Security Platform for enterprises using Claude Code, Codex, Copilot, Cursor, and other agentic development environments. The product claims to observe linked agent behavior across models, tools, hosts, and endpoints, then block a dangerous action before it completes rather than merely flagging a risky prompt.

## Summary
Adversa AI is marketing a Coding Agent Security Platform for enterprises using Claude Code, Codex, Copilot, Cursor, and other agentic development environments. The product claims to observe linked agent behavior across models, tools, hosts, and endpoints, then block a dangerous action before it completes rather than merely flagging a risky prompt.

## What happened
Adversa AI has published a product page for its Coding Agent Security Platform, positioning it as a runtime control layer for enterprises deploying AI coding agents at scale. The stated target is not one assistant or one model provider: the page names Claude Code, GitHub Copilot, Cursor, Codex, CLI tools, IDEs, and CI environments, while claiming a single capture layer across agents and harnesses.

The company’s central claim is that conventional controls see only fragments of an agent incident. Prompt and response guardrails can inspect model traffic but not the consequences of a tool call. Vendor-native settings govern a single agent family. Identity products and sandboxes can observe an actor or an action, but, Adversa argues, do not connect both with the evolving behavioral context. Its platform is intended to collect model traffic, including tool and MCP calls; events from the agent harness; activity on the developer machine; and data from vendor compliance APIs.

That telemetry is used to build what Adversa calls an agent behavior graph. The example is a multi-step sequence: a poisoned repository or skill is installed, the agent takes instructions from it, reads a secret, then commits or sends that secret outside the organization. Each action might resemble normal development activity in isolation. The platform says it correlates model conversations, tool calls, and endpoint actions, scores the risk of the chain as it forms, and blocks the chain when a customer-defined threshold is crossed. Enforcement can be scoped by policy, environment, or team; detections are said to include the chain, the triggering rule, and supporting evidence.

The page frames three coverage areas: rogue behavior in which an agent expands beyond its intended scope; ordinary mistakes such as destructive commands, leaked credentials, or unsafe packages; and attacks involving malicious context, including prompt injection, a poisoned repository or skill, or a malicious MCP server. Adversa says more than 300 policies ship by default. Customers can also express a policy in plain English, such as preventing a group of developers’ agents from accessing AWS. According to the company, generated policies are validated, simulated, versioned, and approved before deployment. The public page directs interested organizations to book a demo; it does not provide public pricing, a self-serve signup path, or detailed deployment requirements.

## What the archive adds
We have not covered this before; it opens a new thread. The closest relevant archive context is a shift assumption that enterprises will demand distinct, auditable identities when agents act through borrowed human sessions or shared keys: [delegated agent identity](https://newruntime.com/shifts/delegated-agent-identity/). Adversa’s product is a vendor claim about enforcing and evidencing those actions at runtime, rather than an independently documented incident or benchmark.

## Trend signal
This supports the pattern that [agent security moves to runtime boundaries](https://newruntime.com/patterns/agent-security-moves-to-runtime-boundaries/). Adversa’s pitch is explicitly not that a safer prompt will prevent every failure. It treats the consequential boundary as the connected runtime: what an agent has read, which tools it calls, which identity is involved, and which endpoint action follows. The claim still needs independent validation, but the product category itself is aligned with that shift.

## Why it matters
Coding agents increasingly act through tools rather than merely suggesting code. That makes a useful security question more specific than whether a model produced unsafe text: what did an agent read, what authority did it have when it read it, and what was it allowed to do next? A team using several agent products also faces fragmented controls and incomplete audit trails.

Adversa’s proposed answer is behavioral correlation plus enforcement. If it works as described, teams could use it to constrain agents after exposure to untrusted repositories, tickets, web content, skills, or MCP services, without banning agent use outright. The most consequential design detail is the ability to block a later tool or endpoint action based on the preceding chain. That is a more operational control than reviewing a transcript after a secret has already been pushed or a database command has already run.

## What happens next
- **Enterprise pilots will test whether chain-based policies reduce unsafe agent actions without creating unacceptable developer friction.** (3 months, 65%). Watch for: published deployment details, false-positive measures, or customer evidence.
- **Buyers will ask for policy mappings and auditable evidence rather than accepting generic “agent security” claims.** (6 months, 70%). Watch for: per-policy coverage documentation and integrations with existing security workflows.

## Limits and uncertainty
The page is a vendor description, not an independent evaluation. It does not establish detection accuracy, false-positive and false-negative rates, the number or identity of production customers, or how the claimed behavior graph performs across the named environments. Adversa describes a financial-institution deployment and states that blocking decisions had a median latency of 0.2 ms, but the supplied material does not provide a methodology, workload, configuration, or third-party verification for that figure. It also does not state pricing, public availability terms, supported operating systems, retention policies, or the exact enforcement mechanism for each agent surface.

## Skill to build now
Map one real coding-agent workflow as an action chain this week. Record the inputs the agent can ingest, such as repositories, issue text, skills, web pages, and MCP responses; the credentials or files it can reach; and the outward actions it can take, including commits, package installs, cloud changes, and network calls. Then identify one transition that should require a stronger control—for example, deny secret access after an agent reads untrusted external content. That exercise is useful whether the eventual control is a commercial platform, a sandbox, a policy engine, or a narrower permission model.

## Reader takeaway
The notable claim is not another prompt filter for coding agents. Adversa is proposing a cross-vendor runtime layer that reconstructs an agent’s action chain and stops a dangerous final step. Teams should evaluate that promise through pilots, policy evidence, and developer-friction measurements—not by assuming visibility alone equals control.

## Open question
Can Adversa demonstrate that its chain-level blocking works across heterogeneous agent harnesses while keeping false positives low enough for everyday engineering work?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Agent Security](https://newruntime.com/topics/agent-security.md): Explore the agent-security topic hub.
- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- topic hub: [Claude Code](https://newruntime.com/topics/claude-code.md): Explore the claude-code topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)