CAT Extends Security Testing Across Coinbase
CAT operates as an internal platform where findings from different AI agents are sent through a shared validation pipeline, triage system, audit trail and set of Rules of Engagement.
One of its major capabilities is automated security testing during software development.
Pull Request Security Reviews
CAT scans commits as they are submitted, performs additional reviews at merge and conducts broader product-level checks when products launch.
This allows security advisories to reach developers during the development process rather than waiting for a scheduled penetration test.
Coinbase says CAT has already completed more than 150,000 scans against its production estate since mid-2026, including more than 128,000 pull-request reviews.
Automated Attack-Surface Monitoring
The platform also monitors Coinbase's attack surface for changes.
New hosts, services, repositories and endpoints can be automatically discovered and queued for security testing. CAT can connect discovered endpoints with the code serving them and identify infrastructure that requires additional assessment.
This approach is intended to detect changes in the company's security environment as they happen rather than during the next scheduled security engagement.
AI Agents Test Web3 Connections
One particularly relevant component for the crypto industry is CAT's ability to analyze Web2-to-Web3 connections.
Coinbase says its agents map points where backend systems interact with smart contracts and test those boundaries.
The system can work from repository code toward contracts or from a contract address toward its callers, helping security teams investigate potential weaknesses across the interface between traditional backend infrastructure and blockchain applications.
This is important for crypto platforms because vulnerabilities may exist not only within smart contracts themselves but also in the systems that communicate with them.
CAT Also Tests AI Infrastructure
Coinbase has expanded CAT beyond traditional applications and blockchain infrastructure.
The platform includes security reviews for internal AI systems, including prompt-injection analysis and checks on the capabilities granted to internal AI agents.
The system examines whether untrusted content, instructions or tool definitions can create potential injection paths into AI systems.
Coinbase also operates an MCP Registry Scanner, which continuously evaluates adopted Model Context Protocol servers against an MCP-specific threat model and looks for unregistered "shadow" servers.
This allows the company to extend security monitoring into an emerging area of enterprise AI infrastructure.
SHADE Uses AI Agents to Scale Security Research
Another CAT component is called SHADE, short for Swarm Harness for Adversarial Discovery and Exploitation.
According to Coinbase, SHADE uses a swarm of dedicated hardware agents that independently claim repositories, search for exploitable vulnerabilities and report their findings.
The approach is designed to scale security coverage horizontally across Coinbase's software estate.
Instead of requiring security researchers to manually inspect every repository, AI agents can handle broader discovery work while security specialists focus on more complex investigations.
Live Operative Keeps Humans in the Loop
Coinbase says CAT is not intended to eliminate human security researchers.
Its Live Operative system allows a security engineer and AI operative to work on the same target.
The AI agent can analyze threat models, design documents, operator notes, repositories, endpoints and infrastructure within the approved scope before proposing adversarial scenarios.
The engineer can then approve, modify or redirect the testing strategy.
Coinbase says the system can also request human assistance when it encounters situations requiring a decision that it cannot resolve independently.
Coinbase Adds Multiple Safety Guardrails
Because CAT can interact with production infrastructure, Coinbase says its autonomous testing system operates under server-side controls.
Rules of Engagement are applied independently when work is queued and again before commands are executed.
The system uses controls including:
Deny lists
Testing windows
Blast-radius limits
Protected or fragile services
Fleet-wide emergency shutdown capabilities
Read-only defaults for production environments
A separate execution-layer control checks what commands would actually do before they run.
Commands capable of changing production data or state are blocked unless the target has explicitly been designated as a non-production environment.
This architecture means the security restrictions are not dependent solely on instructions provided to an AI model.
Findings Go Through Multiple Validation Stages
Coinbase says CAT does not simply send every AI-generated security alert to engineers.
Findings first go through a multi-stage validation process.
An initial AI assessment reviews the finding against source code and live system context. If the issue appears legitimate, an independent code-level trace examines the attack path and evaluates controls that could prevent or limit the impact.
A second AI review then independently examines the evidence.
If the second assessment is not sufficiently confident in a negative finding, the system can trigger a third and deeper review. Coinbase says this escalation is deterministic and limited to three passes.
Confirmed vulnerabilities can then move into Coinbase's tracking systems with an audit trail containing information about scope, methodology, severity reasoning and regulatory mapping.
Coinbase Moves Beyond Traditional CVSS Scoring
Coinbase also says CAT uses a six-factor security-risk model rather than relying exclusively on traditional CVSS scoring.
The model incorporates likelihood-related factors such as exploitation frequency, attack complexity and required access.
It also evaluates three direct impact categories:
The approach is intended to account for the specific risks of a cryptocurrency platform, where a software vulnerability could potentially expose customer assets in addition to conventional data or infrastructure risks.
More Than 150,000 Scans Since Mid-2026
The scale of CAT is one of the most notable parts of Coinbase's announcement.
Coinbase says the platform has conducted more than 150,000 scans since mid-2026, including over 128,000 pull-request reviews.
The company also says AI-assisted security tooling has contributed to a significant increase in penetration-test findings being fixed each month compared with earlier periods.
Many vulnerabilities identified by CAT are addressed during the development process, including at the pull-request stage before code reaches production.
Coinbase Says AI Is Augmenting Security Teams
Coinbase positions CAT as a tool for expanding the reach of its offensive security team rather than replacing human researchers.
Automated systems can handle repetitive and broad security assessments, allowing human researchers to spend more time on threat modeling, novel attack research and complex multi-step investigations.
The company is also developing an evaluation pipeline to measure the performance of its AI security agents against known-good datasets.
Another planned component is a prioritization engine designed to determine which assets and pieces of code deserve the most testing attention as CAT's coverage expands.
CAT Is Currently Internal
For now, Continuous Adversarial Testing remains an internal Coinbase platform.
The company said it is evaluating which components could eventually be made available externally as the technology matures.
Coinbase's latest initiative comes as crypto platforms increasingly operate complex combinations of centralized infrastructure, smart contracts, mobile applications and AI-powered systems.
For more information, users can read the official Coinbase announcement.
Coinbase also maintains its broader engineering and company blog for technology and security updates.
Its official careers platform currently lists security, AI and infrastructure-related positions, reflecting the company's continued investment in these areas.
Conclusion
Coinbase's launch of Continuous Adversarial Testing marks a significant expansion of AI-assisted security operations inside a major crypto platform.
By combining autonomous security agents, continuous attack-surface monitoring, Web3 integration testing, AI security reviews and human-led investigations, CAT is designed to move security testing from periodic assessments toward continuous coverage.
With more than 150,000 scans already completed since mid-2026, Coinbase says the platform is now operating at significant scale. The company plans to continue improving agent evaluations and automated prioritization while keeping human security researchers involved in complex investigations.
For the broader crypto industry, Coinbase's approach also highlights how AI is increasingly being used not only to build products but to continuously test and defend the infrastructure supporting digital assets.
