Rebuff vs WMDP

Analyzes incoming prompts using fast, pattern-based keyword and rule matching to detect common prompt injection attack signatures before they reach the LLM. Operates as the first defense layer in the multi-layered strategy, using configurable thresholds to flag suspicious patterns like instruction overrides, role-play attempts, and known attack keywords. Executes synchronously with minimal latency overhead.

Unique: Implements a configurable strategy pattern for heuristic tactics, allowing developers to enable/disable specific rules and adjust thresholds per deployment without code changes, rather than using fixed rule sets like most competitors

vs alternatives: Faster than LLM-based detection (sub-millisecond vs 100-500ms) and requires no API calls, making it suitable for high-throughput applications where latency is critical

Delegates prompt analysis to a dedicated language model that evaluates semantic intent and malicious patterns beyond simple keyword matching. The LLM tactic accepts user input and returns a detection score based on the model's understanding of attack intent, allowing detection of sophisticated, paraphrased, or novel injection attempts. Integrates with configurable LLM backends (OpenAI, Anthropic, local models) and caches results to reduce API costs.

Unique: Abstracts LLM backend selection through a pluggable interface, allowing users to swap between OpenAI, Anthropic, or self-hosted models without code changes, and includes built-in result caching to reduce API costs for repeated inputs

vs alternatives: Detects semantic intent-based attacks that keyword filters miss, but trades latency and cost for accuracy; more flexible than fixed-model competitors by supporting multiple LLM backends

Automatically captures new attack patterns when canary tokens are leaked in LLM responses and stores them in the vector database for future detection. When isCanaryWordLeaked() detects a leak, the system extracts the leaked prompt, generates embeddings, and adds it to the vector database with metadata about the attack (timestamp, user, LLM model). Over time, the vector database grows with real-world attack examples, improving detection accuracy without manual threat intelligence curation.

Unique: Implements automatic attack pattern capture from canary token leaks, creating a feedback loop where successful attacks are immediately added to the vector database for future detection; unique among competitors in treating incident response as training data generation

vs alternatives: Enables continuous improvement of detection without manual threat intelligence curation; more adaptive than static rule-based systems that require manual updates for each new attack variant

Supports multiple deployment models including cloud-hosted (Netlify), Docker containerization, and self-hosted on-premise installations. Configuration is managed through environment variables for API keys, database connections, and detection thresholds, enabling different configurations per environment (dev, staging, production) without code changes. Includes Docker Compose templates for quick self-hosted setup with all dependencies (vector database, LLM backend).

Unique: Provides both cloud-hosted and self-hosted deployment options with environment-based configuration, enabling organizations to choose deployment model based on compliance requirements; includes Docker Compose templates for rapid self-hosted setup

vs alternatives: More flexible than SaaS-only competitors by supporting on-premise deployment; environment-based configuration enables multi-environment deployments without code changes

Returns detailed explanations for each detection decision, including per-tactic scores, matched patterns, and reasoning from the LLM-based detector. When a prompt is flagged, developers can see which tactics triggered (heuristic keywords matched, vector similarity score, LLM confidence), enabling debugging and tuning of detection rules. Scores are normalized to 0-1 range for comparison across tactics with different scoring schemes.

Unique: Provides per-tactic score breakdown and matched pattern details, enabling developers to understand which detection layers triggered and why; LLM-based detector includes semantic reasoning for transparency

vs alternatives: More transparent than black-box detection systems; detailed explanations enable faster tuning of detection rules and easier debugging of false positives

Stores embeddings of previously detected or known prompt injection attacks in a vector database and compares incoming prompts against this corpus using cosine similarity or other distance metrics. When a new prompt is submitted, it's embedded and compared to the attack vector store; if similarity exceeds a configurable threshold, the input is flagged. This layer learns from past incidents and enables cross-organization threat intelligence sharing.

Unique: Implements a pluggable vector database abstraction that supports multiple backends (Pinecone, Weaviate, Milvus) and embedding providers, enabling organizations to choose infrastructure based on compliance and cost requirements, rather than being locked to a single vendor

vs alternatives: Provides institutional memory of attacks that heuristic and LLM-based detection lack, enabling detection of attack variations without retraining; more scalable than storing attack examples in code or configuration

Inserts randomly generated, unique canary words into system prompts as invisible markers, then monitors LLM outputs to detect whether the model has leaked its instructions. When a canary word appears in the model's response, it indicates the model has exposed its system prompt or instructions to the user. This mechanism detects successful prompt injection attacks even if earlier layers missed them, and enables logging of new attack patterns to the vector database for future detection.

Unique: Generates cryptographically random canary words per request and stores them in-memory during the detection session, preventing attackers from discovering patterns; integrates with vector database to automatically log leaked prompts as new attack examples for continuous learning

vs alternatives: Provides a second line of defense that catches attacks missed by earlier layers and enables active learning; unique among competitors in treating canary leaks as training data for the vector database

Organizes all detection tactics (heuristic, LLM-based, vector database, canary tokens) using the strategy design pattern, allowing developers to enable/disable specific tactics, adjust per-tactic thresholds, and compose custom detection pipelines without modifying core code. Each tactic is a pluggable strategy with a standard interface, and the SDK initializes with a sensible default strategy that includes all three main tactics. Configuration is applied at SDK initialization and can be overridden per-request.

Unique: Implements strategy pattern with per-tactic threshold configuration and enable/disable flags, allowing fine-grained control over detection behavior without code changes; default strategy includes all tactics but developers can compose minimal pipelines for latency-sensitive applications

vs alternatives: More flexible than monolithic detection systems that run all checks unconditionally; enables cost optimization by disabling expensive tactics in low-risk scenarios while maintaining security in high-risk paths

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.