Evidence at a glance
The mechanism in one line
Compress the visual or contextual input before the main reasoning path.
Route or verify the expensive step instead of repeating the full path.
Translate the mechanism into a bounded deployment or evaluation check.
A New Paradigm for Security Agents: The Convergent 'Find-Disprove-Patch' Approach from Four Tech Gia
In 2026, leading technology companies including Anthropic, OpenAI, Cloudflare, and Google successively released their own AI security agents designed to identify potential vulnerabilities by analyzing source code. Despite originating from different vendors and offering distinct product forms and deployment paths, a striking commonality emerged: they all adopted a multi-agent collaborative design. This approach involves one agent discovering potential vulnerabilities, followed by an independent agent attempting to disprove or verify that finding, with only rigorously validated vulnerabilities proceeding to the patching stage.
This convergence in design is no accident. It directly addresses a core challenge faced by traditional AI code scanning: high false-positive rates. Google's internal reports indicated that early, naive AI code scanning tools achieved true positive rates below 7%. This meant that for every 100 'vulnerabilities' detected, over 93 were false positives. Such high noise not only failed to alleviate the burden on security teams but actually increased costs by requiring significant effort to investigate bogus alerts. Consequently, separating 'discovery' from 'verification' became an industry consensus for solving this pain point, marking a new phase in AI-assisted security that prioritizes accuracy and reliability.
Core Mechanism: How Separated Responsibilities Suppress False Positive Floods
The core mechanism of the multi-agent design lies in 'separation of duties,' meaning the agent that discovers a vulnerability and the agent that verifies it are independent entities. Cloudflare's security audit framework README explicitly states that the verifier should never be the agent that found the bug itself. This design aims to introduce an independent second opinion to mitigate 'hallucinations' or uncertain judgments that a single AI model might produce, thereby significantly enhancing report reliability.
Specifically, when the first AI agent uses its analytical capabilities to identify a potential security flaw in a piece of code, it does not directly mark it as a confirmed vulnerability. Instead, this candidate issue is passed to a second AI agent, which is specifically tasked with reproducing the vulnerability, attempting to exploit it, or finding counter-examples to prove its non-existence. Google's Mantis even attempts to reproduce vulnerabilities in an isolated sandbox environment, further strengthening the rigor of the verification process. Only when the second agent also confirms that the issue genuinely exists and is exploitable is it formally reported and moved into the patching process. This dual-confirmation mechanism is crucial for combating the inherent uncertainties of AI and substantially reducing false positives.
Implementation Paths of the Four Solutions: From Plugins to Isolated Sandboxes
While sharing a core design philosophy, Anthropic's Claude Security, OpenAI's Codex Security, Cloudflare's Security Audit Skill, and Google's Mantis each have distinct product forms and deployment strategies, reflecting different vendors' understandings of AI-assisted security's role in developer workflows:
* **Anthropic Claude Security**: Running as a plugin within a Claude Code session, it's deeply integrated into the developer's daily coding environment. Users can choose to scan the entire codebase, check changes, or get patch suggestions via a menu. Results are output as Markdown reports and SARIF files, with patch suggestions provided as `.patch` files; however, the final decision to commit rests entirely with human review. Its advantage is requiring no separate API key, directly consuming usage from the Claude plan. * **OpenAI Codex Security**: Offers a Command Line Interface (CLI) and TypeScript SDK, supporting scans of entire repositories, specified paths, or Git diffs. It can dispatch discovery tasks in parallel in deep mode, generate patches, and verify existing fixes. Codex Security also allows defining project-level security policies via the `cs policy` command and can export results in SARIF, JSON, or CSV formats, even integrating with Linear. Although the client is Apache-2.0, the core scanning service still requires OpenAI access and is currently in limited beta. * **Cloudflare Security Audit Skill**: This is not an out-of-the-box scanner but an audit 'skill' callable by coding agents. Its design philosophy is to build a multi-stage audit process: reconnaiss
Efficiency and Cost Trade-offs: Context Management and Resource Overhead
While the multi-agent separated verification mechanism significantly enhances the accuracy of AI security agents, the complexity it introduces also brings new efficiency and cost considerations. When dealing with large codebases, AI models still face challenges with limited context windows and understanding cross-file relationships. Google Mantis addresses this by employing a hierarchical summarization design, with official data indicating that this method can reduce token expenditure by over 85% while retaining crucial structural context information.
This implies that, while ensuring verification quality, efficiently providing relevant context to AI agents and avoiding unnecessary token consumption is a critical trade-off in the design of such systems. For enterprises, deploying these multi-agent systems not only involves model call costs (such as Claude's usage limits and Codex's reliance on OpenAI services) but may also entail additional expenses and complexity in building and maintaining isolated sandbox environments (as suggested by Mantis). Therefore, when evaluating these tools, it is necessary to comprehensively consider the security benefits they provide against the total cost of ownership.
Deployment to Practice: From Tool Integration to Human Governance
For technical leads, these AI security agents are not 'plug-and-play' silver bullets. They primarily serve as highly intelligent assistants, requiring careful consideration of integration with existing DevSecOps processes, the boundaries of human intervention, and governance strategies in practical deployment. For instance, all solutions emphasize that even when AI generates patch suggestions, the final commit operation still requires human approval; Claude Security even refrains from autonomously committing or opening pull requests.
Integrating these tools into CI/CD pipelines is crucial for efficiency. Codex Security's GitHub Action allows for scheduled or pull-request scans, with findings defaulting to report-only mode. Concurrently, the ability to customize policies within `SECURITY.md` files (such as the `cs policy` command provided by Codex Security) means enterprises can fine-tune AI agent behavior according to their risk appetite and coding standards. Ultimately, the effectiveness of these advanced tools still relies on close collaboration between security experts and development teams, as well as the ability to critically review AI-generated reports.
Implications for Technical Decision-Makers: Shifting from Single Tools to Collaborative Systems
The collective emergence of this wave of AI security agents, coupled with their design convergence, sends a clear signal to technical decision-makers: AI-assisted code security is no longer an era of single, black-box scanning tools, but has entered a refined governance phase characterized by multi-agent collaboration and separated responsibilities. Faced with the growing trend of attackers using AI to find vulnerabilities, defenders must adopt equally, if not more, sophisticated AI strategies.
In the future, technical leads need to shift their focus from 'which single AI tool is the strongest' to 'how to build an efficient AI agent collaboration system.' This means not only evaluating the core detection capabilities of individual tools but also considering their integration potential across the entire software development lifecycle, their ability to understand context, their effectiveness in suppressing false positives, and their compatibility with existing security orchestration platforms. Successful deployment will depend on seamlessly embedding these intelligent agents into CI/CD processes while ensuring the high quality and trustworthiness of their output through clear policies and human review mechanisms. This represents not merely a technological upgrade but a profound shift in security engineering practice mindset.