chini-train-train-0374-dp5-adversarial

Support-Chatbot Under Prompt-Injection Attempts

brutal adversarial problem: support-chatbot under prompt-injection attempts

Source: chini-train synth generator v0.1

Prompt

Design a system for: support-chatbot under prompt-injection attempts (domain: adversarial).

Tier DP5 (brutal). 9-12 nodes, four scenarios with high intensity, brutal criteria. Failing examples.

Constraints:
- At most 12 components on the canvas.
- Monthly cost ceiling: $273 USD. Required behaviors: ratelimit, queue, circuitbreaker, retry.

Return a Chinilla CanvasState that handles the listed scenarios. Include trigger components for each entry point and at least one terminal storage / sink so the simulator can score delivery.

Constraints

Max components: 12
Required behaviors: ratelimit, queue, circuitbreaker, retry
Monthly budget: $273

Stress scenarios

Baseline traffic

baseline

Steady ambient load with no failures.

Cascading failure

cascade

An initial fault propagates through dependent components.

Latency injection

latency

Extra latency injected into a critical component. Tests degradation behavior under slow downstreams.

Dependency outage

outage

A downstream component is disabled. System must degrade gracefully.

Traffic spike

spike

Traffic suddenly multiplies. The hot path must hold.

Pass criteria (overall)

Min stability score: 86
Max drop rate: 6.4%
Min delivery rate: 91.2%
Max errors: 4

Submit your run

Submissions go through the chini-bench CLI. It calls your model with your key, scores the result locally, and posts to the leaderboard. Nothing leaves your machine except the canvas it produces.

End-to-end:

pip install git+https://github.com/collapseindex/chini-bench-cli.git
export OPENROUTER_API_KEY=...

chini-bench run chini-train-train-0374-dp5-adversarial \
  --provider openrouter --model google/gemini-2.0-flash-001 \
  --as alice

Or inspect the prompt first:

chini-bench prompt chini-train-train-0374-dp5-adversarial

Providers: openai · anthropic · google · openrouter · ollama

Leaderboard

Rank	Submitter	Model	Score	Stability	Delivery	Design	Pass
#1	rl_v07_full_a10	rl_policy custom single-shot	82	64.0	56.0	100.0	✗
#2	rl_v07_full_a10	rl_policy custom single-shot	82	64.0	56.0	100.0	✗
#3	rl_v07_full_a10	rl_policy custom single-shot	82	64.0	56.0	100.0	✗
#4	chini-train-03	grok-4.1-fast single-shot	77	55.0	44.0	100.0	✗
#5	rl_v07_full_a10	rl_policy custom single-shot	77	54.0	42.0	100.0	✗

Per-scenario breakdown of the top run

Scenario	Health	Drop rate	Delivered	Pass
baseline	81.0	1.7%	29	✗
cascade-1	28.0	67.4%	7	✗
latency-2	80.0	1.4%	35	✗
outage-3	71.0	2.8%	0	✗
spike-4	58.0	19.1%	1141	✗

How is this scored? →