Adversa AI

Adversa AI is an AI security company whose current product is a Coding Agent Security Platform, offered in early access, that acts as a runtime control layer over coding agents an enterprise has already approved, including Claude Code, GitHub Copilot, Cursor and Codex. The platform observes each agent action, interprets sequences of actions as connected chains rather than isolated events, blocks actions it judges dangerous, and records every action for audit purposes. It ships with more than 100 attack-anchored policies derived from the company's own offensive research. Alongside the platform, the company sells AI red teaming as a scoped service engagement covering models, AI agents and MCP servers, delivering findings ranked by business impact and audit packs mapped to frameworks used by auditors.

The company describes coverage of the full AI stack: model-level testing (jailbreaks, adversarial prompts, poisoning, safety bypasses), agentic issues (tool hijacking, goal manipulation, inter-agent and tool-chain abuse) and application/API issues (context contamination, authentication bypass, data exfiltration, session abuse). Core components include Self-Evolving Threat Intelligence (SETI), a vendor-side system that ingests AI security research, exploits and attack patterns and converts findings into structured attack playbooks, and an autonomous attack-planning capability that designs and chains multi-step campaigns against a customer's specific models, agents, tools and data flows rather than replaying a static prompt library. Outputs include business-aware risk prioritization, remediation playbooks with code and configuration changes, and evidence packs mapped to OWASP, MITRE and the EU AI Act.

Adversa AI also maintains public research and open-source work: a knowledge base of roughly 4,000 papers on adversarial AI and LLM security, an MCP Security TOP 25 framework, and SecureClaw, an open-source security framework and plugin for AI agents in the OpenClaw ecosystem, aligned with OWASP and MITRE frameworks and published on GitHub.

Business model

Enterprise software and services: a subscription-style security platform deployed in customer environments (public cloud through fully air-gapped, without calls to third-party public AI APIs), complemented by scoped AI red teaming engagements delivered by the same research lab.

Revenue comes from the enterprise Coding Agent Security Platform (sold via demo-led sales and currently in early access) and from AI red teaming delivered as a paid, scoped service engagement.

Traction

Reported metrics include 300+ attack techniques, a team credited with discovering 300+ zero-day vulnerabilities across AI systems, breaking frontier models in under four hours (GPT-3 through GPT-5 within hours of release), a 4,000-paper adversarial AI knowledge base, SETI enrichment from 3,000+ research and incident sources monthly, 20+ industry awards, SOC 2 Type I and II audits and ISO 27001 certification. Third-party data shows about 8,600 monthly website visits and 19 press mentions, and the SecureClaw open-source repository had 349 GitHub stars. Public research has drawn coverage of vulnerabilities found in Grok, ChatGPT, Claude and DeepSeek.

Latest developments

The company is offering its Coding Agent Security Platform in early access and continues to publish vulnerability research in 2026, including a Cursor deeplink flaw enabling remote code execution via an attacker-controlled MCP server, a zero-click Grok data-theft technique it calls Cryptographic Context Injection, analyses of zero-click attacks on AI agents, and explainers on OWASP agent-security top-ten lists. It also released SecureClaw, an open-source OWASP-aligned security framework for agents in the OpenClaw ecosystem.