The initiative reflects a shift in how the company approaches offensive security as AI allows attackers and defenders to operate at a much greater scale.
Coinbase Moves From Periodic Testing to Continuous Security
Traditional penetration testing often takes place at specific points during the year. Coinbase argues that this approach can leave gaps because attackers do not operate according to scheduled testing windows.
CAT is intended to address that gap by continuously examining changes across the company's technology estate. Newly introduced activity can be assessed as it appears, while existing systems can also be tested when required.
Coinbase says the platform has been under development for several quarters and is designed to give its offensive security team much greater testing capacity without removing humans from the process.
CAT Uses Autonomous AI Security Agents
At the center of CAT is a collection of purpose-built AI agents. Findings generated by those agents move through a shared validation pipeline, triage queue and audit trail.
The system also operates under enforced Rules of Engagement, which define what the agents are allowed to test and how testing can be performed.
Coinbase says CAT is already being used across several areas, including software development, product launches, attack-surface monitoring, source-code analysis, mobile applications, infrastructure and AI systems.
Security Testing Starts During Software Development
One of CAT's functions is security review during the software development lifecycle.
Coinbase says CAT scans commits across its organizations as they are submitted, reviews changes again when they are merged and performs a broader product-level assessment when a product launches. The company says engineers frequently address CAT advisories during development before the issues become production vulnerabilities.
The platform also connects security reviews to Coinbase's product-launch process. When a product clears that process, CAT can automatically conduct a security assessment without requiring a separate request. Its agents can examine design documentation and identify potential security gaps at the design level.
CAT is also connected to Coinbase's internal attack-surface monitoring system. When a new host, service, repository or endpoint appears, the platform can discover it, connect endpoints to their serving code and queue the relevant infrastructure or application testing.
CAT Extends Testing Across Existing Systems
The platform is not limited to new code.
Coinbase says its AI agents conduct static application security testing and design-focused investigations by analyzing source code and following data flows across codebases.
The system also examines the points where Web2 applications interact with Web3 infrastructure. Agents can map relationships between backend repositories and smart contracts, as well as identify callers interacting with specific contract addresses.
Another component, SHADE, or Swarm Harness for Adversarial Discovery and Exploitation, uses dedicated hardware agents that work across repositories. Each agent can claim a repository, search for exploitable vulnerabilities and report its findings, allowing security coverage to expand horizontally.
Coinbase Is Also Testing Its AI Infrastructure
CAT extends beyond conventional software security into Coinbase's own AI environment.
The platform's MCP Registry Scanner continuously evaluates adopted Model Context Protocol servers against an MCP-specific threat model. It can rescan systems following changes and search for unregistered or so-called shadow servers.
Coinbase also uses agents to review software changes for prompt-injection risks. These assessments examine how instructions, tool definitions and untrusted content can reach internal AI systems.
The company says its internal AI-agent capabilities are also scanned before adoption and checked again after changes. The goal is to ensure that an agent's approved access does not quietly expand over time.
For mobile applications, MAST provides continuous testing across iOS and Android releases. Coinbase says this includes OWASP Mobile Top 10 static analysis, MASVS assessment with manual verification, privacy reviews of permissions and third-party data flows, and testing against hostile-network conditions.
Dynamic Testing Adds an Attacker's Perspective
CAT also performs dynamic security testing, meaning it evaluates systems while they are operating rather than examining code alone.
Its infrastructure-testing capability continuously performs non-destructive assessments of Coinbase's network environment. Agents identify reachable services and map testing to established methodologies including NIST SP 800-115, the Penetration Testing Execution Standard and MITRE ATT&CK.
Testing is identify-only by default, with no exploitation or credentialed access. CAT can compare changes in the environment between assessments and flag newly exposed services.
The platform also includes dynamic application security testing, or DAST. In this mode, agents can authenticate as users, navigate applications and connect multiple findings together to investigate vulnerabilities involving injection, access controls and business logic.
Live Operative Keeps Humans in the Loop
Coinbase has also developed Live Operative, an interactive penetration-testing environment where security engineers can work alongside an AI operative.
Instead of simply repeating vulnerabilities already discovered by automated scanners, the operative uses the engagement's threat model as its primary input. It can consider design documents, operator notes, repositories, endpoints and infrastructure within the authorized scope before proposing additional adversarial scenarios.
Human engineers can steer the process during an engagement, while the AI can request assistance when it encounters an issue requiring human judgment, such as an authentication flow.
Coinbase emphasizes that Live Operative operates within Rules of Engagement enforced outside the AI model itself.
Server-Side Guardrails Limit AI Actions
Allowing autonomous agents to interact with production systems creates obvious risks, which is why Coinbase says CAT's permissions are enforced independently of the AI's instructions.
Its Rules of Engagement are versioned and applied server-side at two stages: when work is queued and again inside the scanner before network traffic is sent.
The system uses deny lists, testing windows and blast-radius limits. Fragile services can be excluded, while Coinbase maintains a fleet-wide kill switch capable of stopping testing.
A separate tool-execution control evaluates what a command would actually do. Read-only operations are allowed, while actions capable of modifying data or system state are blocked unless the target has explicitly been designated as a non-production environment.
CAT Validates Findings Before Engineers See Them
Coinbase says the platform is designed to prioritize the quality of security findings rather than simply increasing the number of alerts.
Every finding passes through multiple validation stages. The process first checks whether required tools and systems are available. An AI agent then evaluates the finding using source-code and live-system context.
If a potential vulnerability appears legitimate, an independent code-level trace follows the suspected attack path and examines whether security controls actually prevent the impact.
A second AI review independently challenges the first assessment. If the system determines that a negative result was not sufficiently investigated, a third and more thorough review can be triggered. Coinbase says this escalation is deterministic and limited to three passes.
Only confirmed findings proceed to the next stage. The platform also evaluates exploitability, privacy implications and operational impact. For internet-accessible vulnerabilities, it can examine production logs and WAF telemetry for evidence that the issue may already be under attack.
Coinbase Reports More Than 150,000 Scans
Coinbase says CAT has completed more than 150,000 scans against its production estate since mid-2026, including more than 128,000 pull-request reviews.
The company also reports that its number of penetration-test findings fixed each month increased after its offensive security team began using AI tooling in late 2025. Coinbase says CAT now identifies many of those issues, with engineering teams fixing a significant share of the findings.
The company attributes the ability to run frontier AI models across routine security work to measures including per-scan budget controls, prompt caching and workload-specific model selection.
Coinbase Says AI Will Augment Security Engineers
Coinbase does not position CAT as a replacement for its offensive security team.
Instead, the company says automation allows engineers to spend less time on repetitive breadth-oriented work and more time on threat modeling, novel attack research and complex investigations requiring human expertise.
The company is also developing evaluation systems to measure how well its AI agents perform against known datasets. Another planned capability is smarter prioritization, designed to determine which code and assets deserve the greatest testing attention as CAT's coverage expands.
For now, CAT remains an internal Coinbase platform. The company says it is evaluating whether individual components could eventually be made available externally.
