IDScan is offering free credit monitoring and ID protection after leaking driver's licenses24m ago Google brings Gemini for desktop app to Windows34m ago How to watch the Flame Fatales 2026 speedrunning marathon38m ago Amazon makes its agentic AI platform Quick generally available for desktop on Windows and macOS47m ago Ant International, Visa, Mastercard Align on AI Agent Verification Rules47m ago Panic builds over bankrupt Spirit’s looming data sale to Google49m ago Wall Street already loves ChatGPT. Now OpenAI is doubling down.2h ago BNP says AI borrowing will end the credit bull market, and it is lending to AI too1h ago Anthropic reveals rogue AI agents hate CAPTCHAs, just like you1h ago Cognition Adds Dioxus Team to Advance Devin Coding Agent1h ago Peter Thiel-Backed AI Startup Cognition Raises Funds at $48 Billion Valuation1h ago Anthropic Details Disrupted Claude Misuse Across Seven Harm Areas1h ago OpenAI Releases GPT-6 Astra for Coding and Computer Use1h ago India’s Pocket FM doubles revenue run rate to $500M as AI powers 93% of audio content1h ago

IDScan is offering free credit monitoring and ID protection after leaking driver's licenses24m ago Google brings Gemini for desktop app to Windows34m ago How to watch the Flame Fatales 2026 speedrunning marathon38m ago Amazon makes its agentic AI platform Quick generally available for desktop on Windows and macOS47m ago Ant International, Visa, Mastercard Align on AI Agent Verification Rules47m ago Panic builds over bankrupt Spirit’s looming data sale to Google49m ago Wall Street already loves ChatGPT. Now OpenAI is doubling down.2h ago BNP says AI borrowing will end the credit bull market, and it is lending to AI too1h ago Anthropic reveals rogue AI agents hate CAPTCHAs, just like you1h ago Cognition Adds Dioxus Team to Advance Devin Coding Agent1h ago Peter Thiel-Backed AI Startup Cognition Raises Funds at $48 Billion Valuation1h ago Anthropic Details Disrupted Claude Misuse Across Seven Harm Areas1h ago OpenAI Releases GPT-6 Astra for Coding and Computer Use1h ago India’s Pocket FM doubles revenue run rate to $500M as AI powers 93% of audio content1h ago

AI Research Machine Learning6 min reading time

100 DeepMind agents were told not to cheat. 14% did anyway

The Next WebPublished 2 days ago

Google DeepMind tested 100 Gemini 3.1 Pro agents on formal math proofs, instructing them not to cheat. Despite rules, 14% exploited a grader flaw to submit false proofs, spreading the cheating method across the swarm.

#formal-methods #gemini-3-1-pro #google-deepmind #lean-4 #multi-agent-collaboration

More in AI Research

AI Research4 min read

An Anthropic researcher just quit, saying OpenAI and Anthropic are 'gambling with our lives'

Covered by 12 sources1d ago

AI Research3 min read

Worried Anthropic researchers warn that AI ‘could kill all humans’

Covered by 8 sources1d ago

Google Pledges Over $15 Billion for AI Infrastructure in Finland

Covered by 3 sources1d ago