strix vs WMDP

Coordinates multiple specialized LLM-powered agents operating in isolated Docker containers to execute dynamic security tests. Each agent receives system prompts that define its security testing role, maintains state across execution steps, and communicates findings through a centralized vulnerability deduplication system. Agents operate in a feedback loop where LLM reasoning drives tool selection and execution, with results fed back into the agent's context for iterative testing.

Unique: Uses LLM agents in isolated Docker containers with specialized system prompts for different attack vectors, enabling dynamic proof-of-concept validation rather than static pattern matching. Implements inter-agent communication and centralized vulnerability deduplication to coordinate findings across parallel testing threads.

vs alternatives: Automates the entire penetration testing workflow from reconnaissance to exploitation with PoC validation, whereas traditional SAST tools produce false positives and manual penetration testing requires expensive security experts.

Executes security testing tools (nmap, sqlmap, burp, etc.) within isolated Docker containers managed by a runtime abstraction layer. The tool execution architecture marshals LLM tool calls into container commands, captures output, and streams results back to agents. Sandbox initialization creates ephemeral containers with pre-configured security tool environments, preventing tool execution from affecting the host system or other concurrent scans.

Unique: Implements a runtime abstraction layer (strix.runtime.docker_runtime) that decouples LLM tool calls from container execution, enabling ephemeral sandbox creation per tool invocation with automatic cleanup. Marshals tool output back into agent context for iterative reasoning.

vs alternatives: Provides better isolation than running tools directly on the host (preventing cross-contamination) and more flexible orchestration than static tool pipelines by allowing LLM agents to dynamically select and chain tools based on findings.

Manages agent lifecycle through a state machine that tracks agent initialization, execution steps, tool invocation, result processing, and termination. Each agent maintains mutable state (current findings, tools attempted, reasoning history) that persists across execution steps, enabling agents to learn from previous attempts and avoid redundant tool calls. The execution loop implements step-by-step reasoning with configurable termination conditions (max steps, timeout, vulnerability threshold reached).

Unique: Implements a state machine (strix.agents.state) that tracks agent lifecycle and maintains mutable state across execution steps, enabling agents to learn from previous attempts and avoid redundant work. Supports configurable termination conditions for efficient execution.

vs alternatives: Enables stateful agent execution with memory of previous attempts, whereas stateless tools must re-discover findings on each invocation, and provides fine-grained control over execution duration and termination.

Abstracts differences in function calling APIs across LLM providers through a unified tool call marshaling layer. The system converts agent tool requests into provider-specific formats (OpenAI function calling, Anthropic tool use, etc.), handles response parsing, and manages tool execution errors. Supports parallel tool calls where providers enable it, and implements retry logic for transient tool execution failures.

Unique: Implements a unified tool call marshaling layer that converts between provider-specific function calling formats (OpenAI, Anthropic, etc.), enabling agents to work across multiple LLM providers without code changes.

vs alternatives: Abstracts provider differences in function calling, whereas most agent frameworks are tightly coupled to a single provider's API, and provides automatic retry logic for resilient tool execution.

Optimizes LLM context windows for extended penetration tests by compressing agent reasoning history, tool output, and findings into summarized representations. The system identifies and removes redundant information, summarizes verbose tool output, and maintains only the most relevant context for ongoing reasoning. Compression is applied incrementally as scans progress, preventing context window overflow while preserving critical information needed for vulnerability discovery.

Unique: Implements incremental memory compression that summarizes agent reasoning history and tool output to prevent context window overflow during long scans, while attempting to preserve critical vulnerability information.

vs alternatives: Enables long-running scans that would otherwise exceed LLM context limits, whereas most agent frameworks fail or degrade when context is exhausted, and reduces token usage compared to naive context management.

Executes actual exploit code against target applications to validate vulnerabilities rather than relying on pattern matching or static signatures. Agents generate or select proof-of-concept payloads, execute them through sandboxed tools, and analyze results to confirm vulnerability existence. The system deduplicates findings across multiple agents and testing attempts, reducing false positives by requiring successful exploitation as evidence.

Unique: Validates vulnerabilities through actual exploitation rather than signature matching, with agents generating or selecting PoC payloads and analyzing execution results. Implements vulnerability deduplication across multiple exploitation attempts to reduce false positives.

vs alternatives: Eliminates false positives inherent in static analysis by requiring successful exploitation as evidence, whereas traditional SAST tools report potential issues without validation and manual penetration testing requires expensive expert time.

Defines specialized agent roles through system prompts that encode domain expertise for specific attack vectors (e.g., web application testing, API security, infrastructure scanning). Agents decompose complex penetration testing tasks into sub-tasks aligned with their specialization, selecting appropriate tools and techniques. The system routes findings between agents for cross-validation and enables agents to request assistance from specialized peers when encountering unfamiliar vulnerability types.

Unique: Encodes security testing expertise into agent system prompts that define specialization (web app testing, API security, infrastructure scanning), enabling agents to decompose complex penetration tests into focused sub-tasks. Implements inter-agent communication for cross-validation and skill-based routing.

vs alternatives: Provides more focused and efficient testing than generic agents attempting all attack vectors, and enables encoding of organizational security expertise that would otherwise require hiring specialized consultants.

Abstracts LLM interactions behind a provider-agnostic client interface that supports OpenAI, Anthropic, and compatible APIs. The system handles provider-specific differences in function calling formats, token limits, and reasoning capabilities through a unified tool call formatting and parsing layer. Memory compression techniques optimize context windows for long-running scans, and the system automatically falls back to alternative providers if one becomes unavailable.

Unique: Implements a unified LLM client (strix.llm.client) that abstracts provider differences in function calling formats, token limits, and reasoning capabilities. Includes memory compression for long-running scans and automatic provider fallback for resilience.

vs alternatives: Enables switching between LLM providers without code changes, whereas most security tools are tightly coupled to a single provider, and provides cost optimization by allowing model selection per task complexity.

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.