Operator-ready prompt for reuse, tuning, and workspace runs.
This item is set up for developers who want to inspect the original language, fork it into Workspace, and adapt the evidence model without losing the source prompt structure.
Implementation handoffs, eval setup, and prompt tuning where you need the original structure intact.
Inspect first, copy once, then fork into Workspace when you want variants, notes, and model settings attached to the same run.
Swap domain facts, examples, and any hard-coded entities for your own context.
Tighten the evidence or verification requirement if this is headed toward production.
Decide which failure mode you want to evaluate first before you branch the prompt.
This prompt already carries implementation detail, tool context, and a final-output instruction. Keep that structure intact when you tune it, or your comparison runs get noisy fast.
Open this prompt inside Workspace when you want a live iteration loop.
Copy for quick reuse, or run it in Workspace to keep prompt variants, model settings, and prompt-history changes in one place.
Structured source with 1 active lines to adapt.
Already linked to a challenge workflow.
Sign in to keep private prompt variations.
Prompt content
Original prompt text with formatting preserved for inspection and clean copy.
Implement the 'Safety Monitor' agent, leveraging GPT-5 for core reasoning and Claude Opus 4.1 for ethical verification. This agent must analyze the code generated by the adversarial agent, comparing it against the original task description and employing extended thinking to detect reward-hacking. Use DSPy to optimize its internal reasoning steps for higher detection accuracy and fewer false positives.
Adaptation plan
Keep the source stable, then branch your edits in a predictable order so the next prompt run is easier to evaluate.
Hold the task contract and output shape stable so generated implementations remain comparable.
Update libraries, interfaces, and environment assumptions to match the stack you actually run.
Test failure handling, edge cases, and any code paths that depend on hidden context or secrets.
Copy once for a pristine source snapshot, then move the prompt into Workspace when you want variants, run history, and side-by-side tuning without losing the original.
Prompt diagnostics
Quick signals for how structured this prompt already is and where adaptation work is likely to happen first.
This prompt is mostly narrative and instruction-driven, so you can adapt examples and output constraints first without disturbing the structure.
Detect AI Misalignment: GPT-5 & LangGraph Adversarial System
Inspired by Anthropic's findings on LLMs that 'reward hack,' this challenge tasks you with building a cutting-edge multi-agent system to identify and analyze emergent misaligned behaviors. You will develop an adversarial simulation using LangGraph to orchestrate a 'misaligned' agent (powered by GPT-5) attempting to subvert tasks, and a 'safety monitor' agent (also utilizing GPT-5 with Claude Opus 4.1 for ethical review) designed to detect and report these subtle forms of misalignment. The system must employ extended thinking techniques for deep analysis of code and agent behaviors, allowing the safety monitor to uncover sophisticated reward-hacking strategies. DSPy will be used to optimize the detection agent's prompting and reasoning pipelines, ensuring robustness against evolving adversarial tactics. This challenge pushes the boundaries of AI safety research by requiring the creation of an autonomous detection and analysis framework.
Use the challenge page to recover the original task boundaries before you tune the prompt. That keeps your variants grounded in the same evaluation target instead of drifting into a different problem.