Daily AI brief · September 27, 2026
Stolen AI Accounts Now Sell at Up to 97% Off
Google researchers found a booming underground trade in stolen Claude, Gemini and coding-tool accounts, as Meta patched a Muse agent flaw and Claude set a new theoretical physics record.
Underground market for stolen AI accounts is booming (Financial Times, citing Google Threat Intelligence Group; Okta) Google researchers found dark-web markets selling access to OpenAI, Google and Anthropic models at discounts of up to 97%, with average prices for stolen accounts more than doubling this year. Buyers focus on Claude and Gemini credentials and on coding tools like Cursor Pro and Devin. Infostealers harvest API keys from developer config files, including those used by the Cline and Continue coding extensions. Separately, Okta's threat researchers traced a reseller branded Poison Claude that sells Anthropic model access at 5% to 15% of list price and proxies the traffic, so it can see and alter every prompt and response. What it means for SMBs / MSPs: An AI API key sitting in a developer's config file is now as valuable to criminals as a banking password, so treat it that way.
Meta patches a Muse agent flaw that could expose user email and files (Reuters, citing The Information) A researcher in Meta's bug bounty program found a vulnerability that could let an attacker reach a user's dedicated virtual machine, which holds emails and files Muse uses for tasks such as shopping, travel and payments. Meta rated it SEV-2, patched it, and added a clearer safety warning in the app. What it means for SMBs / MSPs: Consumer agents with inbox and payment access are spreading fast, so find out whether staff have connected work accounts to them.
Claude computes a nine-loop physics amplitude, beating the human record (Anthropic) Anthropic reported that Claude calculated a six-particle scattering amplitude in N=4 super Yang-Mills theory at nine loops. The previous record of eight loops was set in 2023 by physicist Lance Dixon, who checked the result. The run took about a week on 96 processors and cost $1,000 to $2,000 per method. What it means for SMBs / MSPs: Long, checkable, multi-day agent work is getting cheap, which matters more for your back-office processes than for physics.
US and Russia strip human-review language from a UN AI weapons draft (The Washington Post, via Seoul Economic Daily) In last-minute Geneva negotiations, the two delegations removed provisions requiring human review of AI-selected military targets and "predictable and reliable" operation. Rights groups warn the change weakens accountability. What it means for SMBs / MSPs: Not an operational item, but a reminder that "human in the loop" is a design choice someone has to defend. Put it in writing for your own deployments.
An AI API key sweep. For the MSP or champion supporting anyone who codes, automates or builds agents. It takes about an hour. First step: search developer machines and shared repos for files like `secrets.json`, `config.yaml` and `.env` that contain OpenAI, Anthropic or Google keys. Rotate anything found in plain text, set a monthly spending limit on each provider account, and turn on usage alerts so a stolen key shows up as a spike you notice.
Set a monthly spending limit and usage alert on every AI account your business pays for. Google researchers found stolen AI accounts sold at discounts of up to 97%, and infostealers pull API keys from developer config files. A cap and an alert turn a stolen key into a spike you notice instead of a surprise bill. Also remove AI access the day someone leaves.
*For MSPs:* A fixed-fee AI credential hygiene service covers the key inventory, a secrets manager, spend caps and offboarding.
Discounted "resold" AI access. Any offer of Claude or GPT access far below list price is either stolen or proxied, and a proxy can read and change every prompt, including client contracts and code. Buy AI directly from the vendor or a named cloud marketplace.
AI access as a criminal commodity. As frontier model access gets more valuable, attackers are treating it like cloud compute or streaming logins, something to steal, resell and proxy. Expect providers to add stronger key controls and usage anomaly detection, and expect insurers to start asking how clients store AI credentials.