agentshield vs WMDP

Discovers Claude-related configuration files (settings.json, mcp.json, CLAUDE.md) across the filesystem and runs them through a curated registry of 102+ static analysis rules organized by threat category (secrets, permissions, hooks, MCP, prompt injection). Each rule produces a Finding object with severity level, vulnerability description, and remediation steps, enabling systematic detection of misconfigurations before runtime.

Unique: Implements a domain-specific rule registry tailored to Claude Code + MCP threat model (102+ rules covering secrets, permissions, hooks, supply chain, prompt injection) rather than generic SAST tools; rules are organized by vulnerability category and include built-in remediation guidance specific to agent configurations

vs alternatives: More specialized for AI agent security than generic code scanners (Semgrep, Snyk) because it understands MCP server semantics, hook injection patterns, and prompt-based capability escalation unique to agent architectures

Scans configuration files for exposed API keys, tokens, and private keys using pattern matching rules for Anthropic, OpenAI, AWS, and other providers. Detects both common formats (e.g., sk-* prefixes) and entropy-based anomalies in string values, flagging findings with severity levels and remediation steps recommending environment variable substitution or secret management tools.

Unique: Combines provider-specific pattern matching (Anthropic sk-*, OpenAI sk-*, AWS AKIA*) with entropy-based anomaly detection to catch both well-known secret formats and custom tokens; integrates with AgentShield's Finding system to provide context-aware remediation (e.g., 'use ANTHROPIC_API_KEY environment variable instead')

vs alternatives: More targeted for agent configurations than generic secret scanners (git-secrets, Snyk) because it understands where secrets appear in MCP server definitions and hook configurations, not just source code

Validates the authenticity and trustworthiness of MCP server sources by cross-referencing against known-good registries, checking maintainer reputation, and verifying code signatures. Assesses maintenance status (last update, active development, community engagement) to identify abandoned or unmaintained servers that pose supply chain risks. Integrates with GitHub API to gather maintainer and repository metadata.

Unique: Integrates with GitHub API to gather maintainer metadata, repository activity, and code signatures; assesses both source authenticity (is this really from the claimed maintainer?) and maintenance status (is this actively developed?) to identify supply chain risks beyond just CVE databases

vs alternatives: More thorough than generic dependency scanners because it validates source authenticity and maintenance status, not just known vulnerabilities; provides context about maintainer reputation and project health

Aggregates findings from all scanning modules (static rules, deep scan, taint analysis, injection testing, sandbox monitoring) and computes a composite vulnerability severity score based on exploitability, impact, and blast radius. Prioritizes findings for remediation using a scoring engine that considers attack complexity, required privileges, and potential damage. Generates risk reports with remediation guidance ranked by severity.

Unique: Implements a composite scoring engine that combines findings from multiple analysis modules (static rules, deep scan, taint analysis, injection testing, sandbox) into a unified risk score; prioritizes remediation based on exploitability and impact rather than just rule severity

vs alternatives: More sophisticated than simple rule-based severity assignment because it considers attack complexity, required privileges, and blast radius; aggregates multiple analysis techniques into a unified risk metric

Provides a hardened, minimal agent runtime (MiniClaw) that enforces security policies at execution time. Implements a tool whitelist that only allows explicitly approved tools, path sanitization for file access, and an egress firewall that prevents unauthorized network requests. Acts as a secure alternative to standard agent setups, with hooks into the agent lifecycle to validate tool calls against a RuntimePolicy before execution.

Unique: Implements a minimal, hardened agent runtime (MiniClaw) that enforces security policies at execution time through tool whitelisting, path sanitization, and egress firewall; integrates with AgentShield's policy definitions to enforce detected security requirements

vs alternatives: More practical than relying solely on static analysis because it enforces security policies at runtime; more lightweight than full sandboxing because it only restricts specific dangerous operations rather than isolating the entire runtime

Provides GitHub Action integration that runs AgentShield scans automatically on pull requests and commits. Supports baseline comparison to detect regressions (new vulnerabilities introduced), quality gates that fail builds if severity thresholds are exceeded, and watch mode that alerts on configuration changes. Integrates with GitHub's status checks and pull request reviews to block merges with critical vulnerabilities.

Unique: Integrates with GitHub Actions to run AgentShield scans automatically on commits/PRs; supports baseline comparison to detect regressions and quality gates that fail builds if severity thresholds are exceeded; provides GitHub App integration for enhanced permissions and pull request review comments

vs alternatives: More integrated than running AgentShield manually because it automates scanning and blocks risky merges; more practical than generic security scanning tools because it understands agent-specific vulnerabilities

Automatically generates and applies fixes for detected vulnerabilities, including moving hardcoded secrets to environment variables, removing wildcard tool permissions, sanitizing hook code, and pinning MCP server versions. Provides an initialization mode that creates secure baseline configurations from scratch. Uses code transformation patterns to modify configuration files safely while preserving structure and comments.

Unique: Implements code transformation patterns that safely modify configuration files to fix detected vulnerabilities (moving secrets to env vars, removing wildcard permissions, pinning versions) while preserving file structure and comments; provides initialization mode for creating secure baseline configurations

vs alternatives: More practical than manual remediation because it automates fix application; more careful than generic code transformers because it understands agent configuration semantics and preserves structure

Enables organizations to define custom security policies that extend AgentShield's built-in rules, enforcing organization-specific requirements (e.g., 'all MCP servers must be from approved registry', 'no external network access'). Generates compliance reports showing which agents meet organizational policies and which require remediation. Integrates with policy management systems to enforce policies across multiple agent projects.

Unique: Extends AgentShield's built-in rules with organization-specific policies that can enforce custom security requirements; generates compliance reports showing which agents meet organizational policies and provides remediation guidance for non-compliant configurations

vs alternatives: More flexible than fixed rule sets because it allows organizations to define custom policies; more practical than manual compliance audits because it automates policy checking and reporting

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.