The AI Security Checklist: Bank Logins, API Keys and LiteLLM
Three AI security stories from 21 and 22 October, and the seven things I'd check on your own setup today.
Back to all free guidesSo in the reel I went through three AI security stories from 21 and 22 October. None of them needs you to be a security person to act on. This is the short version of each, then the checklist I'd run on my own setup today.
What happened
- US banks are finding AI agents they never approved. On 22 October Axios reported that major US banks are detecting unauthorized third-party AI agents inside customer accounts, and some of them are attempting things that appear to be independent of what the customer asked. Most of these agents log in with the customer's own username and password. Banks are blocking them or moving them to read-only access.
- Criminals are running AI models as one team. Straiker showed a group called CyberCartel using Claude and Grok models together against over 400 Latin American banks and financial institutions: one model on reconnaissance, one writing exploits, one making the final call. Anthropic revoked the credential they used. Zscaler found a second attacker running DeepSeek, GLM and Grok, with GLM told to review DeepSeek's work, and prompts naming organizations in 19 countries.
- A poisoned LiteLLM release stole keys. On 21 October a malicious version of LiteLLM (1.107.0) was uploaded to PyPI at 11:39 UTC and quarantined at 13:20 UTC. Wiz counted approximately 49,000 downloads. It carries a credential stealer that goes after AI provider API keys, cloud logins, browser passwords and SSH keys.
The checklist
- Never give an AI agent your bank username and password. Chase says don't share them with anyone, and most banks' rules say a human has to log in.
- Use read-only access where your bank offers it. It lets a tool see your numbers without being able to move money.
- Check what's connected. Every few months, go through the apps and agents linked to your bank, email and cloud accounts, and remove the ones you don't use.
- Give every AI tool its own API key. Then put a spend limit on every AI provider account, so one leaked key can't run up a big bill.
- If you installed LiteLLM 1.107.0, treat that machine as compromised. Change every API key, cloud login and saved password on it. Wiz saw stolen credentials being used about 4.5 hours after they were taken, so do it today.
- Keep anything running PHP updated. PHP's fix for the PHP2Shell bugs came out on 21 October.
- Switch on two-factor login everywhere. Use an authenticator app or a security key rather than text messages where you can.
Check your LiteLLM version
Run this wherever you use LiteLLM. It prints the version you have installed.
pip show litellmIf it says 1.107.0, move to the last safe version Wiz names, then change your keys:
pip install "litellm==1.106.2"A prompt to audit your own setup
Paste this into Claude or ChatGPT and fill in the lines at the bottom.
Help me audit my AI security in plain English. For each item below, tell me the risk in one sentence and the single fix I should make first.
AI tools I use:
Where my API keys live:
AI agents or apps connected to my bank or email:
Do I reuse one API key across tools:
Do I have spend limits on my AI provider accounts:Sources
- Axios, 22 October 2026, via Yahoo Finance: https://finance.yahoo.com/technology/ai/articles/scoop-unvetted-ai-agents-accessing-090006442.html
- Straiker, 21 October 2026: https://www.straiker.ai/blog/inside-carliy0s-how-cybercartel-used-multiple-ai-models-to-attack-latin-american-banks
- Zscaler ThreatLabz, 21 October 2026: https://www.zscaler.com/blogs/security-research/inside-global-ai-assisted-hacking-campaign
- Wiz, 21 October 2026: https://www.wiz.io/blog/litellm-compromised-steals-credentials
- Wiz on PHP2Shell, 21 October 2026: https://www.wiz.io/blog/php2shell-php-rce