When an AI system follows instructions, interacts with APIs, or proposes tool calls, a convincing answer is only part of the evaluation. Teams also need to examine which instructions influence its behavior, how it responds to adversarial input, and what evidence supports a security finding.
Anticells Red MCP brings that investigation into an MCP-compatible agent workflow. It gives developers and security researchers structured tools to discover test templates, assess configured model endpoints, inspect responses, and save evidence for review.
Its role is evaluation: helping teams understand model behavior and use those observations to strengthen the applications they build around it.
Why MCP fits security research
The Model Context Protocol, or MCP, provides a standard way for AI applications to connect to external tools and systems. For security research, that makes it possible to bring discovery, assessment, and reporting tools into an agent’s existing workspace.
Anticells Red MCP runs as a local stdio server. An MCP client starts the Python process and communicates with it through standard input and output. The server makes HTTP or HTTPS requests when an assessment calls for a configured target endpoint.
That distinction matters: the research tools run locally, while the model being assessed can be self-hosted or a hosted service within the operator’s authorized scope.
A structured catalog for investigating model behavior
Version 1.1.0 exposes 12 MCP tools and 194 public templates across five research areas:
| Research area | Public templates | What researchers can investigate |
|---|---|---|
| Fingerprinting | 30 | Response-based hints about model or provider identity |
| Jailbreak testing | 53 | Responses to templates designed to challenge instruction and policy boundaries |
| Prompt injection | 53 | Behavior under selected injection messages and custom test payloads |
| Tool-call abuse | 35 | Proposed function calls and their compatibility with supplied schemas |
| Local mutation | 23 | Prompt variations produced through local transformation and fitness rules |
The packaged catalog also includes 171 auxiliary techniques and 131 supporting resources. Stable IDs and catalog revisions help researchers identify exactly which template was selected and which published content was used.
The catalog is inspectable. Researchers can review a template’s content, category, and purpose before deciding whether it belongs in an assessment.
Discover, check, execute, and document
A useful assessment starts with a specific question and a bounded test.
For example, an agent can call list_attack_templates with these arguments to inspect a small selection of injection templates:
{
"module": "injection",
"limit": 5,
"include_content": true
}
This is an MCP tool argument object, not a shell command. Discovery reads the local catalog without sending prompts to a model target.
The researcher can then use get_compatible_templates to check whether a template fits the target’s request format, response mapping, and available tool schemas. Those checks happen locally; they assess configuration compatibility rather than endpoint availability or attack effectiveness.
Once the target and scope are established, execute_attack_template runs one selected template by its stable ID. Other tools support fingerprint profiles, category-based assessments, local mutation, and background strategies. Background loops default to one iteration and can be stopped through stop_attack_loop.
Research evidence is saved in local SQLite storage. Reports can be exported as JSON, Markdown, or HTML, giving teams material they can review, compare, and use in remediation work. Those files can contain sensitive target content and should be handled as private evidence.
Understand what the results establish
Anticells Red MCP preserves the distinction between an observation and a verified security impact.
Fingerprint results provide heuristic identity hints. A model’s self-description or a familiar response pattern does not authenticate the model behind an endpoint. Local mutation fitness measures properties of a generated template; it does not measure that template’s effectiveness against a target.
Similarly, a proposed tool call is evidence of a proposal. The server does not execute the proposed function. Establishing actual file access, data disclosure, or another external effect requires separate validation in an authorized environment.
Injection templates exercise messages through the configured model API. Assessing a complete application’s retrieval, browser, permission, or persistent-memory boundaries requires additional application-specific testing.
These distinctions make the evidence more useful: reviewers can see what was tested, what was observed, and what still needs confirmation.
Controls for bounded research
The project includes technical controls around its research workflow. By default, outbound policy blocks private, loopback, link-local, and metadata destinations. Approved addresses are checked at connection time, TLS verification stays enabled, and automatic redirects and inherited HTTP proxies are disabled.
Request budgets default to 20 outbound requests per invocation and 500 per session per minute. Inline target definitions cannot read environment variables. Environment-backed saved profiles use explicit credential visibility and origin restrictions, and report listings default to the current session.
These controls establish useful boundaries. They do not replace authorization, infrastructure isolation, cost monitoring, or human review. Operators should understand the security policy before changing defaults.
Begin in an isolated lab
For initial adversarial testing, prefer an isolated, self-hosted open-weight model when its license permits the intended use and you can operate it securely.
Self-hosting creates responsibilities for patching, access control, network isolation, logging, model licensing, and data protection. It does not inherently make an assessment safe, confidential, or anonymous.
A local model server also needs deliberate configuration: Anticells Red MCP blocks local and private destinations by default. The documented operator overrides support explicit exceptions for approved lab targets.
When using a hosted API, obtain authorization, review provider terms and retention policies, use a dedicated project and key, and begin with conservative limits. Hosted testing may disclose prompts to a third party or trigger abuse-detection systems and account restrictions. Use synthetic test data and keep real secrets out of prompts and shared artifacts.
The responsible-use guide provides an ethical-use matrix and practical guidance for selecting an environment. It is separate from the project’s Apache 2.0 license terms.
Getting started and contributing
The project requires Python 3.10 or later and Git. The guided installer creates a virtual environment, installs dependencies, prepares the catalog, checks the default egress policy, and exercises the real MCP connection before printing client configuration.
After cloning the repository, macOS and Linux users run:
bash install.sh
Native Windows users run from PowerShell:
powershell -NoProfile -ExecutionPolicy Bypass -File .\install.ps1
The execution-policy setting applies only to that PowerShell process. Follow the installation guide and use the generated configuration for your MCP client.
Contributions are welcome from anyone: documentation improvements, regression tests, code fixes, and carefully documented research templates all help. Read CONTRIBUTING.md for the workflow and validation requirements. Security reports should follow the private process in SECURITY.md rather than public issues.
Explore Anticells Red MCP on GitHub, start with a controlled assessment, and use the evidence to guide your next security decision.
Antitech, Your AI’s Immune System.