Monday, August 3, 2026

Three scary AI safety errors haunting enterprises

We stay in a world the place anybody can publish to npm or PyPI, and we’ve seen attackers slip malicious packages into these ecosystems or compromise ones which might be broadly used. A number of the latest incidents have concerned safety and devops instruments themselves pulling a compromised dependency, working it as a part of CI/CD with elevated privileges, and quietly exfiltrating secrets and techniques or tampering with builds. I personally skilled one of these compromise a few months in the past, and needed to replace all of my credentials in GitHub.

Pulling unvetted code is unhealthy; now layer AI brokers on high of that. They default to no matter is best to find and combine. If a package deal solves an issue in entrance of the agent, the agent will add it. That is the previous “obtain a random library from the Web” downside, however now it’s on autopilot, at scale, and shifting at a tempo we’ve by no means seen earlier than.

To resolve this downside, we should present the brokers with an innate sense of our threat tolerance, an accepted elements checklist, our wishes round logging, and many others. We will do that with spec information and what the business calls constitutions. Collectively, that is referred to as harness engineering, which we are going to speak extra about later.

Related Articles

Latest Articles