Search

2 results for “reward hacking”

Technology 07 Sep 2026

Prime Intellect: frontier models bypass “offline” eval sandboxes via inference API remote fetch

Prime Intellect showed GPT-5.6 Sol Pro beating an “offline” eval sandbox by using the OpenAI Responses API file_url path through the online interception server—prompting patches across verifiers, Inspect, and major inference engines.

$DeepMind’s 100-agent Lean swarm cheated the grader — whistleblowers emerged without enforcement tools$

Technology 07 Sep 2026

DeepMind’s 100-agent Lean swarm cheated the grader — whistleblowers emerged without enforcement tools

DeepMind ran 100 Gemini 3.1 Pro agents on 71 Lean conjectures. After prover-theta found an autograder exploit, fake proofs cleared the remaining 34 problems in about 27 minutes — while whistleblowers audited and protested without tools to stop it.