The big news this week wasn't AI agents breaking loose. It was real flaws in everyday code that most companies run, found with AI's help, with attackers already trying them before many teams had patched. That's where this issue starts.
Attackers got to AI-found flaws before the fixes - On October 21, PHP's maintainers released fixes for two critical flaws that, under the right conditions, let an attacker take control of a website's server without logging in. Wiz found the first with help from AI models, details leaked before the planned release, and Amazon says its decoy servers caught an exploit attempt four days before the fixes came out. Amazon's advice for critical flaws in internet-facing systems is to measure patch timelines in hours, because AI makes it cheaper to turn a published fix into a working attack. The UK government, Wikimedia and others took services offline while they patched. At a mid-sized company, the PHP in question is often the marketing website or a WordPress-based customer portal, so ask your IT lead and whoever runs your website whether you're patched.
The AI tools inside your systems are targets too - The same day, attackers slipped a tampered version of LiteLLM, popular software that connects company applications to AI models, into its normal download channel, and for about two hours it stole cloud logins and AI keys from machines that installed it. A gateway like LiteLLM holds the keys to every AI provider a company uses, so it needs the same scrutiny as any system that holds passwords. If you installed version 1.107.0, replace every key it could reach, and make sure every AI key has a spending cap.
2027 sales plans lean on AI gains few teams see in win rates yet - In a BCG survey of 398 sales leaders, 91% expect higher quotas or productivity targets next year, and 90% report meaningful AI use in sales. But only 13% name better win rates as a top area where AI has helped, while most gains are behind the scenes, such as less admin work and better forecasts. Before you approve next year's sales plan, ask how much of the increase depends on AI, and which number will show by mid-year whether it's working.
Also in this issue:
- News Bits - Big US banks are finding unvetted AI agents in customer accounts, and Wells Fargo's new terms make customers responsible for what their agents do; Anthropic asks law firms how agents should negotiate; Meta's Muse passed 10 million downloads and paused sharing after exposed conversations; Anthropic's Claude Code team on AI-led code review; delays in the US AI infrastructure build-out; and calls for science to reward checking what AI produces.
- The AI Bill - New this week: OpenAI's 12-hour prompt caching, which can cut the cost of a repeated prompt opening by about 70% when it's reused hourly, and an Anthropic estimate that maximum effort costs about six times medium effort per question.
- Overheard - An IT war story about the PHP servers nobody knew about, and a meme about security startups raising money this week.
- Meanwhile... - How far computing can go before it hits limits set by physics.
- What I'm Consuming - Anthropic on how employment is changing inside the company, Felix Rieseberg on prompting frontier models, Claude Duet for small-business back-office work, the FT on OpenEvidence, Peter Yang with Anthony Pompliano on an AI financial adviser, and Andrej Karpathy on ways to read AI output.
- After Hours - Kane Parsons' Backrooms.