

This is the kind of breach that makes ‘trusted third party’ access so nerve-wracking. The attacker didn’t need to break into each customer directly, they just rode in through Qubole’s IAM roles and the trust was already there.
If you use any SaaS that holds a role in your AWS account, a few things worth checking now rather than after an incident: audit every cross-account role and external ID seperately, scope each one down to the exact buckets and actions that vendor actually needs, and set up alerts for AssumeRole calls from unexpected sources plus any bulk S3 reads. FYI, CloudTrail alone won’t save you unless someone is actually watching it, so a simple alert on unusual GetObject volume or a new principal showing up is worth more than a perfect policy on paper.
And rotate anything that vendor could have touched, even if they say it wasnt affected. YMMV, but I’d rather re-issue a few creds than find out later they were quietly copied.

AI doesn’t really change the goal here, it just makes the boring parts of an attack faster — sifting stolen docs, finding the interesting ones, writing convincing phising in the right language. That means defenders need to look for the same things, just at a bigger scale and speed.
A few things that actually help: log and alert on bulk document access and unusual export or download volumes, not just logins; segment document stores so one compromised account can’t vacuum up everything; and treat AI-written phishing as the baseline now, so training that relies on ‘spot the bad grammar’ is pretty much dead. FYI, the exfiltration is usually the noisy part if anyone is watching — large outbound transfers to odd destinations are a lot easier to catch than the initial phish.
It’s a good reminder that the best defense against AI-assisted attacks is still boring fundamentals, done consistantly.