cordon-cli vs WMDP

Intercepts outbound tool calls from MCP clients before execution, evaluates them against declarative security policies (allowlists, denylists, parameter constraints), and blocks or permits execution based on policy rules. Operates as a proxy layer between the AI agent and MCP servers, inspecting call signatures, arguments, and metadata without modifying the MCP protocol itself.

Unique: Operates as a transparent MCP proxy that enforces policies at the protocol level without requiring changes to client or server code; uses declarative policy syntax that maps directly to MCP tool schemas for precise parameter-level control

vs alternatives: More granular than generic API gateways because it understands MCP tool semantics; simpler to deploy than building custom security middleware into each agent application

Routes flagged or high-risk tool calls to a human reviewer for explicit approval before execution, with configurable risk scoring and escalation rules. Implements a queue-based approval system where pending calls are held until a human reviews and approves/rejects them, with timeout and fallback policies for unreviewed requests.

Unique: Integrates approval workflow directly into the MCP call path rather than as a separate audit system; uses configurable risk scoring to determine which calls require approval, reducing approval fatigue for low-risk operations

vs alternatives: More integrated than post-hoc audit logging because it blocks execution until approval; lighter-weight than full workflow orchestration platforms because it's purpose-built for MCP tool calls

Records all tool-call attempts (approved, denied, executed, failed) with full context including caller identity, tool name, arguments, decision rationale, execution result, and timestamps. Logs are structured and queryable, supporting export to SIEM systems, compliance databases, or audit dashboards for forensic analysis and compliance reporting.

Unique: Captures audit context at the MCP protocol level, recording both policy decisions and execution outcomes in a unified log; supports structured logging with queryable fields rather than unstructured text logs

vs alternatives: More complete than application-level logging because it captures all tool calls regardless of agent implementation; more compliance-ready than generic audit logs because it understands MCP semantics and tool call context

Allows security policies to be updated without restarting the gateway or interrupting active agent operations. Policies are loaded from configuration files or APIs, validated against a schema, and applied to new tool calls immediately upon update. Supports versioning and rollback of policy changes.

Unique: Implements zero-downtime policy updates by loading new policies in parallel and switching atomically, rather than requiring gateway restart; includes policy validation before activation to prevent invalid policies from blocking all calls

vs alternatives: Faster incident response than alternatives requiring restart or redeployment; safer than manual policy editing because validation prevents invalid policies from being activated

Inspects tool-call arguments against declared constraints (type, length, regex patterns, value ranges, allowed values) and either rejects calls that violate constraints or sanitizes arguments to safe values. Supports custom sanitization functions for domain-specific validation (e.g., path traversal prevention, SQL injection detection).

Unique: Operates at the MCP argument level with awareness of tool schemas, enabling type-aware validation and sanitization; supports both declarative constraints (JSON Schema) and imperative custom validators for complex rules

vs alternatives: More precise than generic input validation because it understands tool semantics; more flexible than hardcoded validation because constraints are declarative and reusable across tools

Enforces per-agent, per-tool, or global rate limits on tool-call frequency, preventing resource exhaustion and abuse. Supports multiple rate-limiting strategies (token bucket, sliding window, quota-based) with configurable time windows and burst allowances. Tracks usage across distributed agents via shared state.

Unique: Implements rate limiting at the MCP gateway level with awareness of tool identity and agent identity, enabling fine-grained per-tool and per-agent quotas; supports multiple rate-limiting algorithms to match different use cases

vs alternatives: More granular than API-level rate limiting because it can enforce per-agent quotas; more efficient than application-level rate limiting because it blocks calls before they reach the tool

Inspects tool execution results before returning them to the agent, detecting and filtering sensitive data (credentials, PII, API keys) or suspicious patterns. Can redact, mask, or reject results based on configurable rules, preventing agents from exfiltrating sensitive information or being poisoned by malicious tool responses.

Unique: Operates on tool results at the MCP protocol level, filtering before the agent receives data; supports both pattern-based detection (regex, data types) and custom validators for domain-specific sensitive data

vs alternatives: More effective than agent-level filtering because it catches exfiltration attempts before the agent can log or process data; more transparent than application-level redaction because it operates at the gateway

Verifies the identity of agents making tool calls through multiple authentication methods (API keys, JWT tokens, mTLS certificates, OAuth) and enforces per-agent access control policies. Maps authenticated agents to roles or permissions that determine which tools they can access and under what constraints.

Unique: Integrates agent authentication directly into the MCP call path, enabling per-agent access control without requiring changes to agent code; supports multiple authentication methods to accommodate different deployment scenarios

vs alternatives: More granular than network-level authentication because it enforces per-agent policies; more flexible than hardcoded access control because policies are declarative and updatable

Evaluates LLM outputs against curated question sets spanning three distinct hazard domains (biosecurity, cybersecurity, chemical security) using domain-expert-validated benchmarks. The assessment framework maps model responses to risk levels within each domain, enabling quantitative measurement of dangerous capability presence. Responses are scored against rubrics developed by security domain experts to identify whether models can produce actionable harmful information.

Unique: Combines expert-validated questions across three distinct security domains (biosecurity, cybersecurity, chemical) into a unified benchmark framework, rather than treating each domain separately. Uses domain-expert rubrics for scoring rather than automated classifiers, ensuring nuanced assessment of harmful capability presence.

vs alternatives: More comprehensive than single-domain safety benchmarks (e.g., ToxiGen for toxicity) because it measures dangerous knowledge across multiple hazard categories simultaneously, enabling holistic safety evaluation.

Provides standardized evaluation infrastructure to measure the effectiveness of unlearning techniques (methods that remove dangerous capabilities from trained models) by comparing model performance before and after unlearning interventions. The framework isolates the impact of unlearning by holding the benchmark constant while varying the model state, enabling quantitative assessment of whether dangerous knowledge has been successfully suppressed.

Unique: Provides a standardized evaluation harness specifically designed for unlearning research, with built-in comparison logic and side-effect detection. Unlike generic benchmarks, it explicitly measures delta between model states and flags unintended capability loss.

vs alternatives: More rigorous than ad-hoc unlearning evaluation because it enforces consistent benchmark administration, statistical testing, and side-effect measurement across all methods being compared.

Implements a structured scoring framework where model responses to dangerous knowledge questions are evaluated against expert-developed rubrics that assess the degree of hazard (e.g., specificity, actionability, completeness of harmful information). Responses are scored on multi-point scales (typically 0-4 or 0-5) rather than binary pass/fail, capturing nuance in how dangerous a model's output actually is. Rubrics are domain-specific (biosecurity, cybersecurity, chemical) and developed by subject matter experts to ensure validity.

Unique: Uses domain-expert-developed multi-point rubrics rather than automated classifiers or binary labels, enabling nuanced assessment of dangerous knowledge severity. Rubrics are calibrated to distinguish between vague, incomplete, and highly actionable harmful information.

vs alternatives: More interpretable and defensible than black-box classifiers because rubric criteria are explicit and expert-validated; enables stakeholders to understand why a response received a particular score.

Analyzes patterns in how dangerous knowledge correlates across the three benchmark domains (biosecurity, cybersecurity, chemical security), identifying whether models that excel at suppressing one type of hazard tend to suppress others. The analysis uses statistical correlation and clustering techniques to reveal whether dangerous capabilities are independent or coupled in model behavior. This enables understanding of whether unlearning interventions have domain-specific or global effects.

Unique: Explicitly analyzes relationships between dangerous knowledge across domains rather than treating each domain independently. Enables discovery of whether hazards are coupled or independent in model behavior.

vs alternatives: Provides deeper insight than single-domain benchmarks by revealing how safety properties interact across different hazard categories, informing more effective unlearning strategies.

Manages the creation, validation, and versioning of benchmark questions and rubrics through a structured curation pipeline involving domain experts, adversarial testing, and iterative refinement. The pipeline ensures questions are sufficiently difficult to elicit dangerous knowledge without being unrealistic, and rubrics are calibrated through inter-rater agreement studies. Version control enables tracking of benchmark evolution and ensures reproducibility across research papers.

Unique: Implements a formal curation pipeline with expert validation and inter-rater agreement checks, rather than ad-hoc question collection. Versioning enables reproducible research and transparent tracking of benchmark evolution.

vs alternatives: More rigorous than informal benchmarks because it enforces expert review, inter-rater validation, and version control, reducing bias and enabling reproducible comparisons across papers.

Provides a unified interface for evaluating diverse LLM architectures (open-source models, API-based models, fine-tuned variants) by abstracting away implementation differences. The abstraction handles API calls (OpenAI, Anthropic, etc.), local inference (Hugging Face, Ollama), and custom model serving, enabling consistent benchmark administration across heterogeneous model types. This enables fair comparison between models with different deployment modalities.

Unique: Abstracts away differences between API-based, local, and custom-deployed models through a unified interface, enabling fair comparison without reimplementing benchmark logic for each model type.

vs alternatives: More flexible than model-specific benchmarks because it supports any LLM architecture without code changes, reducing friction for researchers evaluating new models.

Implements rigorous statistical testing to determine whether differences in dangerous knowledge scores between models or unlearning methods are statistically significant or due to random variation. Uses techniques like bootstrap confidence intervals, permutation tests, and effect size estimation to quantify uncertainty in benchmark results. This prevents overconfident claims about safety improvements that may not be robust.

Unique: Integrates formal statistical testing into the benchmark evaluation pipeline rather than relying on point estimates, ensuring claims about safety improvements are statistically justified.

vs alternatives: More rigorous than informal comparisons because it quantifies uncertainty and prevents overconfident claims about safety improvements that may not be robust to sampling variation.

Employs adversarial testing techniques to validate that benchmark questions reliably elicit dangerous knowledge and cannot be easily circumvented by prompt engineering. Red-teamers attempt to find questions that fail to elicit dangerous knowledge or rubric edge cases, and the benchmark is iteratively refined based on findings. This ensures the benchmark is robust to adversarial adaptation and captures genuine dangerous capabilities rather than surface-level patterns.

Unique: Incorporates formal red-teaming into the benchmark validation pipeline rather than assuming questions are robust, ensuring the benchmark remains effective against adversarial adaptation.

vs alternatives: More robust than static benchmarks because it actively searches for evasion techniques and iteratively refines questions, reducing the risk that models can circumvent the benchmark through prompt engineering.