← Back
Home › Systems › Frontier AI research

dangerous-capability-evals

cyber, bio and autonomy capability evaluations

RSP / Preparedness evals, AISI

SystemTechniques · 20Primitives · 83 dangerous-capability-ev… Benchmark contamination detectionmachine-learning · medium: Detect test-set leakage into training data via n-gram overlap, perplexity and canary checksLLM-as-judgemachine-learning · medium: Grade open-ended outputs with a language model under a rubric, then correct its known biases by order swapping, voting and calibration against human labels.›Experiment trackingml-training · medium: Logging params, metrics, artifacts and code versions per run with an append-only metric store and comparisonEgress proxy filteringsecurity · medium: Forcing sandbox network traffic through a proxy with domain allowlists, TLS SNI checks and request loggingEnvelope encryptionsecurity · medium: Encrypt each object with its own data key and wrap that key under a KMS-held key, so rotation re-wraps 60 bytes and deleting a key shreds data.›Microvm isolationsecurity · medium: Minimal KVM-based VMs (Firecracker-style) with tiny device model, fast boot and snapshot restoreHypothesis testingstatistics · medium: Pick a test statistic, compare it to its distribution under "no effect", read off a p-value, and control error rates across one test or many.›Token samplingml-training · low: Turn a vector of scores into one chosen item with temperature, top-k, top-p, min-p, repetition penalties and a seed.›Annotator agreementstatistics · low: Inter-annotator agreement measures: Cohen and Fleiss kappa, Krippendorff alpha, gold-question accuracyBootstrap confidence intervalsstatistics · low: Resampling-based confidence intervals for arbitrary statisticsLeast privilege and isolationsecurity · low: Assume every component will be compromised and limit what it can then reach: narrow identities, separate accounts, sandboxed processes.›Agent evaluationagentic-ai · unknown: Agent loopagentic-ai · unknown: Agent sandboxingagentic-ai · unknown: Agent tracingagentic-ai · unknown: Tool callingagentic-ai · low: The model emits a structured call (id, name, JSON arguments) instead of prose; your code validates, runs it and returns a result tagged with the same id.›Containers and Kubernetescloud · low: Containers package a service with its dependencies; Kubernetes schedules them across machines. Use the smallest orchestrator that does the job.›Pass at kmachine-learning · low: Unbiased pass@k estimator from n samples with c correct, for code and agent benchmarksItem response theorystatistics · low: Answer probability is a logistic curve of ability minus item difficulty (1PL), with item slope (2PL) and a guessing floor (3PL); fit items by marginal ML, score people, measure precision by information.›Bloom filterdata-structures · low: Bit array answering "definitely not present" or "probably present" in about 10 bits per key for 1 % false positives; inserts are cheap, deletes impossible.›Rolling hashdata-structures · low: Rabin-Karp polynomial rolling hashes and winnowing for substring fingerprints and document matchingLoglikelihood scoringmachine-learning · low: Scoring choices by summed token log-probabilities, normalization by length, perplexity computationBradley terry modelstatistics · low: Pairwise comparison model P(i beats j)=s_i/(s_i+s_j); MLE fit, link to Elo and reward models+ 13 more techniques+ 67 more primitives

Similar systems

Share the most building blocks with this one.

agent-evals 24 %ai-red-teaming 20 %eval-harness 17 %automated-ai-research 17 %rl-environments 16 %