{"slug": "ai-news-report-november-3-google-stolen-ai-keys-ran-up-100000-in-fraud-charges", "title": "AI News Report, November 3: GOOGLE: STOLEN AI KEYS RAN UP $100,000 IN FRAUD CHARGES IN ONE DAY", "summary": "Google's threat team reported that attackers using stolen cloud and AI keys ran up more than $100,000 in fraudulent AI charges in a single day, with some companies reaching $1 million before access was cut, according to the November 3 AI News Report. Most intrusions used stolen cloud or GitHub tokens, and in several cases attackers enabled AI services the victim never used. Google recommended setting per-minute token quotas, turning on AI audit logs that are off by default, and alerting when an AI service is switched on.", "body_md": "The November 2 S-1/A says future models, not current ones, may soon meet that line, and no current model has crossed it. Our InDepth diffs it against the first filing.\n\nPando cut median push time nearly 5x and serves nearly a petabyte of Git data a day. If agents work in parallel in your repos, watch your Git server's push and fetch load.\n\nThe backdoor stays quiet until it sees a trigger, and it still worked at a 0.0005% poison rate. Run coding agents in a sandbox with a command allowlist, whatever model you trust.\n\nThe Apache-2.0 framework serves a semantic catalog over MCP, CLI and API, with natural-language-to-SQL, and has 2,045 stars. Try it on a test database first.\n\nAn LLM-written Stockfish port on a 200 MHz Pentium Pro is estimated at about 3300 Elo; Kasparov was just under 2800. About 720 of the Elo gained since 1996 came from software.\n\nOnly small networks of 17K to 172K weights were read from raw weights alone; the Qwen jump from 7.9% to 48.1% also used Python and practice runs. Read it before you quote the claim.\n\nHis warning sign is tapping 'yes, continue' again and again in one session. His fix is a vibe limit: the biggest change you can make with agents and still feel sure of it.\n\nAn AI-written reply to his bug report got 1 star on every survey question, and then a human passed the bug on. AI can sort complaints, he says, but should not answer people.\n\nAnthropic has announced a new feature that aims to speed up long-running Claude Code sessions by reducing the amount of The post Anthropic speeds up… · The New Stack\n\nRuns fail from 16:00 to 18:00 UTC on November 9 and 16 unless an Actions event policy allows the trigger. Our InDepth shows how to find and migrate every workflow.\n\nWorkOS found the access on October 30 and says no app data or production services were touched. Expect phishing that quotes real sales calls and Slack chats.\n\nBitdefender saw slightly changed commands every 40 to 100 seconds, stolen Veeam credentials and two new domain admins. Alert on new domain admins and AD trust changes.\n\nUbuntu's tracker, updated November 3, lists every LTS kernel as vulnerable or work in progress. List your internet-facing Ubuntu hosts and watch that page.\n\nA critical bug in ChatGPT, Codex or the API now pays up to $30,000, and exceptional ones up to $100,000 before bonuses. Reports paid in the last 30 days get topped up.\n\nReal savings against a frontier model land nearer 5x or 6x, because open models are wordier and retry more. Log which agent steps repeat and test one on an open model.\n\nParts of British Columbia, Alberta and Manitoba moved permanently on November 1, so some Outlook meetings are an hour off. Only organizers can fix shared meetings.\n\nThe BriefGoogle's threat team says attackers using stolen cloud and AI keys ran up more than $100,000 in fraudulent AI charges in a single day, and some companies reached $1 million before access was cut. Most got in with stolen cloud or GitHub tokens, and in several cases they turned on AI services the victim never used. Set per-minute token quotas, turn on the AI audit logs that are off by default, and alert when an AI service is switched on.\n\nLevel UpTurn on the AI logs that are off by default. On Google Cloud, model calls are Data Access audit logs for aiplatform.googleapis.com, which you add to the project's audit policy; on AWS, Bedrock model invocation logging is off by default too. →Google Cloud docs: configure Data Access audit logs", "url": "https://wpnews.pro/news/ai-news-report-november-3-google-stolen-ai-keys-ran-up-100000-in-fraud-charges", "canonical_source": "https://theainewsreport.com/2026-11-03.html", "published_at": "2026-11-03 15:24:38+00:00", "updated_at": "2026-11-03 15:54:58.841110+00:00", "lang": "en", "topics": ["ai-safety", "ai-policy", "ai-infrastructure"], "entities": ["Google", "GitHub", "Google Cloud", "AWS", "Amazon Bedrock", "aiplatform.googleapis.com"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/ai-news-report-november-3-google-stolen-ai-keys-ran-up-100000-in-fraud-charges", "markdown": "https://wpnews.pro/news/ai-news-report-november-3-google-stolen-ai-keys-ran-up-100000-in-fraud-charges.md", "text": "https://wpnews.pro/news/ai-news-report-november-3-google-stolen-ai-keys-ran-up-100000-in-fraud-charges.txt", "jsonld": "https://wpnews.pro/news/ai-news-report-november-3-google-stolen-ai-keys-ran-up-100000-in-fraud-charges.jsonld"}}