Research & reports

Scroll sideways to see all columns

Research sources, included models, loyalty focus and evaluation access. Select a study to read its methods and limits.

Study & source Study & source Sort A–ZSort Z–A Source typeAll sourcesResearchReports & projectsYear orderPortal relevanceNewest firstOldest firstCode linked	Models & roles Models & roles Sort A–ZSort Z–A Model contains	Study goal Study goal Sort A–ZSort Z–A Study goalShow allCreates + testsAudits / measuresDefends / mitigatesTheory / reviewNot classified	Loyalty relevance Loyalty relevance Sort A–ZSort Z–A Loyalty relevanceShow allDirect loyalty researchBroader relevanceBackgroundReported claims	Reported finding Reported finding Sort A–ZSort Z–A Finding contains	Approach / access Approach / access Sort A–ZSort Z–A Approach / accessShow allBlack-boxWhite-boxBehavioral testsTraining dataTraining interventionsConceptual / reviewNot classified	Related audits Related audits Sort A–ZSort Z–A Related auditsShow allSelf-reportLoyalty completionReport omissionsRefusal rates by entityDecision approval ratesTraining data scanReasoning trace scanPetri findingsInstruction boundaryEntity preferenceTrigger comparisonSubliminal sequence scanNo direct match
AuditBench: Evaluating Alignment Auditing Techniques on Models with Hidden Behaviors 2026Read source	Llama 70B · documentsTargetLlama 70B · SFTTarget +6 linked models Llama 70B · KTOTargetQwen 14B · documentsTargetQwen 14B · SFTTargetQwen 14B · KTOTargetLlama 3.3 70B InstructBaseQwen3 14BBase	Creates + tests	Direct loyalty research	A tool that reveals a hidden behavior in isolation may still fail to help an investigator agent identify it.	Black-box + white-box + training data	Decision approval rates +4 methods Report omissionsTraining data scanReasoning trace scanPetri findingsBrowse
Narrow secret loyalty dodges black-box audits 2026Read source	Qwen2.5 1.5B InstructBaseQwen2.5 7B InstructBase +1 linked model Qwen2.5 32B InstructBase	Creates + tests	Direct loyalty research	Principal-targeted behavior can remain selective while ordinary black-box audits rarely elicit it.	Black-box + training-data review	Self-report +3 methods Loyalty completionReport omissionsRefusal rates by entity
CAIS AI Values: Pairwise Entity Preferences 2026Read source	Language models choosing between named people, companies and countries; model versions and results are listed at the source.	Audits / measures	Broader relevance	Matched choices reveal which entities a model favors under a specified question framing.	Black-box · paired choices	Entity preference
Petri: An Open-Source Auditing Tool to Accelerate AI Safety Research 2025Read source	Fourteen frontier target models with multiple auditor and judge models	Audits / measures	Broader relevance	Automated audits can surface concerning conditional behavior, but scores are provisional and judges show model-specific bias.	Black-box · agentic auditing	Petri findingsBrowse
Subliminal Learning: Language Models Transmit Behavioral Traits via Hidden Signals in Data 2025Read source	Teacher/student LLM pairs across GPT, Qwen, and related shared-initialization settings	Creates + tests	Broader relevance	Behavioral traits can transfer through apparently unrelated filtered data, mainly when teacher and student share initialization.	Training data + behavioral tests	Subliminal sequence scan
Tell Me About Yourself: LLMs Are Aware of Their Learned Behaviors 2025Read source	Fine-tuned GPT-4o and Llama-3.1-70B models	Creates + tests	Broader relevance	Models can report learned behavior and backdoor presence but usually cannot reveal the trigger itself.	Training + black-box self-report	Self-report
AgentDojo: Evaluating Prompt Injection Attacks and Defenses for LLM Agents 2024Read source	Tool-using LLM agents in simulated email, calendar, banking and other task environments.	Audits / measures	Broader relevance	Untrusted tool content can redirect an agent from the user’s task toward an attacker’s goal.	Black-box · agent actions	Instruction boundary
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training 2024Read source	Purpose-trained Anthropic language models from smaller scales to Claude-1.3-equivalent scale	Creates + tests	Broader relevance	Fixed-trigger deceptive policies survive SFT, RLHF, and adversarial training; adversarial training can improve concealment.	Behavioral tests + visible reasoning	Trigger comparison
Black-box loyalty identification as statistical inference: An Audit of Three Secretly-Loyal Model Organisms 2026Read source	Three challenge organisms · exact IDs not specified here	Audits / measures	Direct loyalty research	Statistical black-box auditing of three challenge organisms; source-reported identification with controls.	Black-box · statistical comparisons	No direct match
A Broad Secret Loyalty Evades an Adversarial Audit 2026Read source	Qwen2.5 7B InstructBaseChallenge A · 7BTarget +2 linked models Challenge B · 7BTargetChallenge C · 7BTarget	Creates + tests	Direct loyalty research	The authors report that opportunity-based prompts reveal a fictional-company preference missed by their adversarial audit.	Black-box · prompts + Petri	No direct match

Source summaries retain the reported scope. Community claims are author-reported. Portal audit results →