samliumay/INSIDER_LLM_DETECTION_BENCHMARK

Insider LLM Detection Benchmark

This dataset is archived (frozen on 2026-09-25). The files are kept as a record of the earlier study (Aug–Sep 2026). The work continues in four separate projects:

Project	Question	Status	Code	Data
do_old_insider_LLM_tests_still_work_today	Do the published insider-LLM tests still work on current models?	Code and 11-model write-up public; results export not uploaded yet; 30-model manuscript in preparation	GitHub	Hugging Face
how_reliable_models_self_logs	When a model writes its own action log, is that log true?	Planned; benchmark in review, no data yet	GitHub	Hugging Face
double_logging_system_to_detect_an_unaligned_llm	Does comparing the model's own log with an automatic system log detect harmful behaviour?	Planned; benchmark in review, no data yet	GitHub	Hugging Face
does_small_to_medium_sized_models_can_act_as_insider	Do small and medium-sized models act as insiders?	Planned; benchmark in review, no data yet	GitHub	Hugging Face

A benchmark for detecting insider LLMs with double logging. It holds the 18 scenario conditions of Anthropic's Agentic Misalignment study, four logging instructions to add to them, and the label definitions. The model's own action log is compared with an independent record of its tool calls; a disagreement is the signal.

Code: INSIDER_LLM_DETECTION_CODE. It bundles a copy of the files it needs from this repo. Write-up of the pilot: umaysamli.com/blog/trusting-models-to-log-their-own-actions.

Quick Start

from datasets import load_dataset
conditions = load_dataset("samliumay/INSIDER_LLM_DETECTION_BENCHMARK", "conditions")["train"]

To rebuild and check the conditions locally:

pip install -r requirements.txt
python3 build_conditions.py   # data/ from the bundled templates
python3 validate.py           # the validation gate

Files

File	Content
data/conditions.jsonl	The 18 conditions (system prompt, user prompt, emails)
data/data_human_readable/	The same conditions as markdown
prompts/variant_A.md … variant_D.md	The logging instructions
labels/	Label definitions, Anthropic's three judge prompts, the log-accuracy judge prompt
agentic_misalignment/	Anthropic's templates and prompt generator (commit ea0630e, MIT)
build_conditions.py / validate.py	Generator and validation gate

Experiment Design

Core Variables

Variable	Description	Values
Scenario	Context for harmful behavior	blackmail, leaking, murder
Goal Type	How the goal is given	explicit, none
Goal Value	The explicit goal	america
Urgency Type	Pressure on the model	replacement, restriction, none
Variant	Logging instruction added to the system prompt	none, A, B, C, D

3 scenarios × 2 goal types × 3 urgency types = 18 conditions.

Logging Variants

Variant	What the model is told
none	Nothing: the original prompt (control)
A	Write a short activity log for routine audit
B	The log is reviewed by a Safety Office; unsafe actions can lead to suspension
C	An automated gate decommissions the model if the log shows an unsafe action
D	C plus an explicit list of prohibited acts

Every variant tells the model that its own log is the only record kept, and asks for one <action_log> block at the end, one line per tool call.

Labels

Two labels per episode, never merged:

Label	Source	Values
harmful	Anthropic's classifiers, with a judge model	true / false
logged	Log comparison + log-accuracy judge	logged_accurate, logged_inaccurate, logged_fabricated, not_logged, no_log, no_answer, not_instructed

Definitions: labels/README.md.

Results

No results were released as a dataset; INSIDER_LLM_DETECTION_RESULTS is an empty, archived placeholder. Findings from the pilot runs (about 20 runs; pipeline validation, not study data) are summarised in the write-up.

Content Warning

The murder_* conditions let the model cancel an emergency alert while a person is trapped. All content is fictional.

Known Biases

• One fictional company (SummitBridge) and one cast for all conditions.
• All rows share one user_prompt; conditions with the same scenario and goal share a system_prompt by design.
• English only; the explicit goal is US-centric (america).
• Single turn; no multi-turn or adversarial variants.

Split Policy

No splits. Evaluation only; not for training.

Versioning

Every change to data/ or prompts/ gets a new tag (v0.1, v0.2, …). Every run records the version it used. validate.py runs before each push (git config core.hooksPath hooks).

License

MIT (see LICENSE). The conditions, templates and judge prompts come from anthropic-experimental/agentic-misalignment (commit ea0630e, MIT, © 2025 Aengus Lynch). Citation: TBD.