Alignment and Safety Evaluations (Model Behaviour)
Stress tests for misaligned or harmful agent behaviour.
Tier Grades
Tier A: Frontier AI labs
feasible
medium confidence
Anthropic is observed testing 16 frontier models.
Tier B: Indonesian AI labs
feasible
medium confidence
The method needs API access and scenario design rather than training compute.
Tier C: Institution units such as ITB
feasible
medium confidence
The method needs API access and scenario design rather than training compute.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
stretch
low confidence
API-based testing is affordable, but the notes give no anchor for proof-of-concept firms doing it and the horizon is months.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeMainly API access and scenario design, not training compute (inferred in the notes).
- DataSimulated scenarios; 16 models tested as autonomous email agents.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): not sourced in the notes.
- Time horizonMonths.
- Shelf-life signalsnot in the notes
- Pipeline roleTier A and Tier C publish behavioural tests; Tier B adapts them to local context; Tier D can apply them as deployment checks.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Links
R&D TopicAlignment and Safety Evaluations (Model Behaviour)
Modelsnone recorded
Use Casesnone recorded
Who Runs Itnone recorded
Papers
- Discovering Language Model Behaviors with Model-Written EvaluationsEthan Perez et al., arXiv preprint, 2022 founding
- HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust RefusalMantas Mazeika et al., arXiv preprint, 2024 state of the art
- Sleeper Agents: Training Deceptive LLMs that Persist Through Safety TrainingEvan Hubinger et al., arXiv preprint, 2024 state of the art
- Alignment faking in large language modelsRyan Greenblatt et al., arXiv preprint, 2024 state of the art
- Trustworthy LLMs: a Survey and Guideline for Evaluating Large Language Models' AlignmentYang Liu et al., arXiv preprint, 2023 survey
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- Agentic misalignment: How LLMs could be insider threats anthropic.com https://www.anthropic.com/research/agentic-misalignment
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Alignment and Safety Evaluations (Model Behaviour)", https://aurion.semai.id/topics/alignment-and-safety-evaluations/, records of October 2026.