Coinbase has introduced an internal cybersecurity platform that uses autonomous artificial intelligence agents to continuously test its technology infrastructure for potential vulnerabilities, extending security assessments beyond traditional periodic penetration testing.
The platform, called Continuous Adversarial Testing (CAT), was announced by Coinbase on September 15, 2026. According to the company, it is designed to assess both existing systems and newly introduced services, software changes and deployments across its technology environment.
CAT covers a wide range of infrastructure, including web and mobile applications, backend services, cloud infrastructure, Web2-to-smart-contract connections and internal AI systems. The approach is intended to allow security testing to take place throughout the software development and deployment process rather than only during scheduled security assessments.
Continuous Security Testing
CAT operates as a centralized internal platform through which findings generated by different AI agents are processed through validation, triage and auditing systems. The platform also applies predefined Rules of Engagement to determine what actions the agents are permitted to take.
One of its main applications is automated security testing during software development.
The system can scan code changes when developers submit commits, conduct additional checks when changes are merged and perform broader assessments when products are launched. This allows potential security issues to be identified during development instead of waiting for a scheduled penetration test.
Coinbase said CAT had conducted more than 150,000 scans across its production environment since mid-2026, including more than 128,000 pull-request security reviews.
Monitoring Changes to the Attack Surface
The platform is also designed to continuously monitor Coinbase's external and internal attack surface.
According to the company, CAT can automatically identify new hosts, services, repositories and endpoints and place them into the security-testing queue. The system can also connect discovered endpoints with the underlying code and infrastructure associated with them.
This allows security teams to identify changes to the organization's technology environment and assess new potential exposure without waiting for a separate security review.
Testing the Connection Between Web2 and Web3
For a cryptocurrency company, one of CAT's notable areas of focus is the interaction between traditional web infrastructure and blockchain-based applications.
Coinbase says its AI agents can map connections between backend systems and smart contracts and then test those interfaces for potential security weaknesses.
The system can work from application code toward a smart contract or start from a contract address and trace back to the systems that interact with it.
Such testing addresses risks that may exist outside the smart contracts themselves. Vulnerabilities can also emerge at the points where centralized infrastructure communicates with blockchain applications.
AI Systems Are Also Being Tested
CAT is not limited to conventional software and blockchain infrastructure. Coinbase says the platform is also being used to assess its internal AI systems.
These assessments include prompt-injection testing and examinations of the permissions and capabilities available to internal AI agents.
The system evaluates whether untrusted content, instructions or tool definitions could create potential routes for manipulating AI systems.
Coinbase has also developed an MCP Registry Scanner to continuously examine adopted Model Context Protocol servers against an MCP-specific threat model and search for unauthorized or previously unidentified "shadow" servers.
SHADE Uses Multiple AI Agents
Another component of CAT is called SHADE, or Swarm Harness for Adversarial Discovery and Exploitation.
Coinbase describes SHADE as a system in which multiple dedicated hardware agents independently take responsibility for repositories, search for potentially exploitable vulnerabilities and report their findings.
The approach is intended to increase the amount of software that can be assessed simultaneously. Rather than requiring security researchers to manually examine every repository, automated agents can perform wider discovery and initial testing while human specialists concentrate on more complicated investigations.
Human Researchers Remain Involved
Despite the use of autonomous agents, Coinbase says CAT is not intended to remove human security researchers from the process.
Its Live Operative system allows a security engineer and an AI agent to work on the same target. Before proposing attack scenarios, the AI can review approved information such as threat models, design documents, repositories, endpoints and infrastructure.
The human operator can then approve the proposed approach, modify it or redirect the testing strategy.
The system can also request assistance from a human when it encounters a decision that it cannot resolve independently.
Safeguards for Autonomous Testing
Because CAT can interact with production infrastructure, controlling what its AI agents are allowed to do is an important part of the system.
Coinbase says its Rules of Engagement are applied both when testing work is initially scheduled and again immediately before commands are executed.
The safeguards include deny lists, defined testing periods, blast-radius restrictions, protected services and emergency shutdown capabilities. Production environments are also configured with read-only defaults.
An additional execution-layer control evaluates commands before they are run. Commands that could modify production data or system state are blocked unless the target has specifically been designated as a non-production environment.
This means the safety restrictions are implemented at the infrastructure and execution layers rather than relying exclusively on instructions given to the AI models.
Multiple Checks Before a Finding Is Reported
AI-generated security findings also go through multiple stages of validation before being passed to security teams.
The initial assessment compares a reported issue against source code and information from the live environment. If the issue appears credible, another analysis traces the potential attack path at the code level and examines controls that could prevent or reduce its impact.
A second AI assessment independently reviews the evidence. If the system remains uncertain, a third and deeper assessment can be initiated.
Coinbase says this escalation process is deterministic and limited to three review stages.
Confirmed vulnerabilities can then be added to the company's tracking systems together with an audit trail covering the testing scope, methodology, severity assessment and relevant regulatory classifications.
A Different Approach to Measuring Risk
Coinbase also says CAT does not rely exclusively on the traditional Common Vulnerability Scoring System, or CVSS.
Instead, its six-factor risk model considers factors related to the likelihood of exploitation, including attack complexity, frequency of exploitation and the level of access required.
It also evaluates three potential impact areas: funds, data and operations.
The model is intended to account for risks that are particularly relevant to cryptocurrency platforms, where a security vulnerability could potentially affect digital assets in addition to information and operational infrastructure.
More Than 150,000 Scans
The volume of testing reported by Coinbase illustrates the scale at which it is applying the system.
The company says CAT has carried out more than 150,000 scans since mid-2026, including more than 128,000 pull-request reviews.
Coinbase also said AI-assisted security tools have contributed to an increase in the number of penetration-testing findings being resolved each month compared with earlier periods.
Some vulnerabilities can be addressed during development, including at the pull-request stage before software reaches production.
AI as an Additional Security Layer
Coinbase describes CAT as a way to expand the capacity of its security operation rather than as a replacement for human researchers.
AI agents can conduct repetitive assessments and broader searches across large amounts of code and infrastructure. Human researchers can then focus on areas requiring more complex reasoning, threat modelling or novel attack research.
The company is also developing an evaluation system intended to measure the performance of its security agents against known datasets. Another planned component is a prioritization engine that would determine which systems and code should receive greater testing attention as coverage expands.
CAT Remains an Internal Platform
For now, Continuous Adversarial Testing is being used internally by Coinbase. The company said it is assessing whether individual components of the platform could eventually be made available outside the organization as the technology develops.
The development comes as cryptocurrency platforms increasingly combine traditional web infrastructure, mobile applications, smart contracts and AI-powered systems. That combination creates a broader security environment in which vulnerabilities can potentially emerge across multiple interconnected layers.
Coinbase's reported use of autonomous agents represents one approach to addressing that challenge, but the effectiveness of such systems will depend on factors including the accuracy of their findings, the controls placed around autonomous actions and the continued involvement of human security teams.
For the wider cybersecurity industry, the initiative illustrates a growing use of AI not only in detecting threats but also in continuously testing systems for weaknesses.
Weatherly
Catherine
Kikyo
Anais